DESC:[0001] The embodiments herein generally relate toappliance-level disaggregation and
prediction of water consumption, specifically, to a system and method for disaggregating and
predicting appliance-level water consumption/usage using IoT devices and artificial
intelligenceto enable non-intrusive, detailed water consumption tracking, enhancing water
10 conservation efforts.
Description of the related art
[0002] In recent years, water conservation has become a critical issue due to the
increasing demand for limited water resources. Traditionally, residential water consumption
is monitored using a single water meter that provides the total water usage for a household.
15 While these systems offer a basic overview of consumption, they do not provide detailed
information on how individual appliances contribute to the total water usage.
[0003] Various solutions, such as manual monitoring using meter readings, smart
water meters, and consumption labels on appliances, have attempted to address this issue.
However, these approaches suffer from several limitations. Manual monitoring is time20
consuming, prone to human error, and provides limited real-time insights. Smart water
meters offer some improvements but often require expensive installations and complex
plumbing modifications. Moreover, appliance consumption labels only provide estimated
usage and may not reflect real-world scenarios, leading to inaccurate data.
[0004] The lack of granular, real-time data on water usage hampers efforts to
25 optimize water consumption and detect leaks or inefficient usage patterns. Therefore, there is
a need for a system that can accurately disaggregate water usage at the appliance level in real
time, without the need for invasive installations or expensive hardware.
[0005] Therefore, there arises a need for a system capable of identifying specific
3
appliances responsible for water usage, predicting future consumption patterns, and
providing actionable insights to users for optimizing water conservation.
SUMMARY
[0006] In view of the foregoing, an embodiment herein provides a system for
appliance-level disaggregation and prediction of water consumption in real-5 time. The system
includes one or more ultrasonic flow sensors, a smart water level monitoring device, a
processor and a remote server. The one or moreultrasonic flow sensors is operatively
mounted on pipelines of respective water-using appliances within a premises and is
configured to measure flow rates of water passing through the pipelines in real-time. The
10 smart water level monitoring device is positioned in a water supply tank and configured to
continuously measure water level in the water supply tank. The processor is
communicatively coupled to the one or more ultrasonic flow sensorsand the smart water level
monitoring device via a network. The processor is configured to preprocess and transmit realtime
sensor data includingflow rates from individual water-using appliances and water level
15 changes in the water supply tank. The remote server is communicatively coupled to the
processor. The remote serverincludes a memory. The remote serverimplements a machine
learning model that is trained using historical labeled flow data. The remote server is
configured toreceivetime-series flow rate data from the pipelines and water level data from
thewater supply tank using the processor.The remote server disaggregatesusing the processor
20 and the machine learning model, a total water usage into individual water-using appliance
usage by analyzing variation in the flow rate and water level changes patterns of water
passing through the pipelines in real-time. The remote server identifies in real-time the
specific water-using appliance responsible for current water usage by applying a trained
appliance classification model to the received flow and water level data. The remote server
25 predicts using a multivariate time series forecasting model comprising a Long Short-Term
Memory (LSTM) neural network, a future water usage for each water-using appliance based
on historical usage patterns and ongoing tank water level changes. The remote
servertransmits the disaggregated appliance-level usage data and the predicted future usage
4
profile to a user device for enabling water conservation actions by a user.
[0007] In some embodiments,the machine learning model is trained using historical
data including a processed dataset of labeled water usage events. The dataset includes flow
rate, velocity, pressure, and water level changes obtained from flow sensors. The flow
sensors provides ground truth data used for calibrating and validating the model 5 to accurately
recognize flow patterns, disaggregate appliance-level consumption, and predict future usage
behaviors. The machine learning model includes at least one of a Long Short-Term Memory
(LSTM) neural network or a one-dimensional convolutional neural network (Conv1D)
configured to model temporal patterns in multivariate water usage data.
10 [0008] In some embodiments, the one or more ultrasonic flow sensors are configured
to generate pulse signals proportional to the flow rate of water, and the processor converts the
pulse signals into digital flow rate data.In some embodiments, the smart water level
monitoring device includes an ultrasonic sensor that is configured to detect minute changes
in water level every two seconds.
15 [0009] In some embodiments, the processor is an IoT edge node configured to buffer
and compress real-time sensor data prior to transmitting to the remote server.In some
embodiments, the remote serveris configured to generate anomaly alerts based on deviations
between predicted and actual usage, indicating potential leaks or inefficiencies.
[0010] In one aspect, a computer-implemented method for appliance-level
20 disaggregation and prediction of water consumption in real-time is provided. The method
includes (i) receiving, using a processor , real-time flow rate data from one or more
ultrasonic flow sensors mounted on pipelines of respective water-using appliances within a
premises, (ii) receiving, using the processor, real-time water level data from a smart water
level monitoring device positioned in a water supply tank, (iii) transmitting, using the
25 processor via a network, the real-time flow rate data and water level data to a remote server
comprising a memory, wherein the remote serverimplements a machine learning model that
is trained using historical labeled flow data, (iv) disaggregating, using the processor and the
machine learning model, a total water usage into individual water-using appliances usage by
5
analyzing variation in the flow rate and water level changes patterns of water passing through
the pipelines in real-time, (v) identifying in real-time the specific water-using appliance
responsible for current water usage by applying a trained appliance classification model to
the received flow and water level data, (vi) predicting, using a multivariate time series
forecasting model comprising a Long Short-Term Memory (LSTM) neural 5 network, a future
water usage for each water-using appliance based on historical usage patterns and ongoing
tank water level changes, and (vii) transmittingto a user device, the disaggregated appliancelevel
usage data and the predicted future usage profile for enabling water conservation
actions by a user.
10 [0011] In some embodiments, predicting future water consumption for each waterusing
appliance includes generating a time-stamped forecast of water usage for each
appliance for a predefined future time interval.In some embodiments, disaggregated water
usage for each appliance is classified into activity categories comprising at least one of:
filling, usage, and no activity, and the machine learning model generates confidence scores
15 associated with each classification.
[0012] In some embodiments, the machine learning model is periodically updated
with newly labeled flow rate and water level data to continuously improve the accuracy of
appliance identification and usage prediction.
[0013] The system enables granular visibility of water usage by disaggregating total
20 consumption into appliance-level usage. Unlike standard household meters that only measure
aggregate water consumption, the system allows a user to identify the exact appliance, such
as a washing machine or geyser, responsible for the water flow at any given time. This level
of detail fosters accountability and enables users to make informed decisions about reducing
excess water usage.The combination of ultrasonic flow sensors on individual appliance
25 pipelines and a smart water level monitoring device in the supply tank provides instantaneous
information on water usage and tank levels. By implementing a Long Short-Term Memory
(LSTM) based forecasting model, the system is capable of predicting future consumption
patterns of each appliance. This predictive functionality allows proactive water management,
6
including scheduling usage, anticipating shortages, and avoiding peak demand conflicts.
[0014] The system also enhances leakage and anomaly detection. Continuous
observation of flow rates and tank water levels enables the identification of leaks, drips, or
abnormal consumption patterns, thereby reducing wastage and minimizing repair or
replacement costs. The machine learning model continuously improves 5 its accuracy by
learning from new usage data, making the system robust against evolving consumption
patterns, seasonal variations, or newly installed appliances.By providing disaggregated usage
data and predictive consumption insights, users can identify inefficient appliances or
excessive consumption behaviors, leading to targeted interventions that reduce overall water
10 demand. This not only conserves water resources but also lowers associated energy costs in
heating and pumping.
[0015] The system provides user empowerment and transparency by transmitting
disaggregated and predicted usage profiles to a user device. This fosters awareness of
consumption behavior and enables households, building managers, or industrial facilities to
15 take timely and targeted conservation actions. Moreover, the wireless connectivity of the
sensors and monitoring devices allows seamless integration into existing smart home or IoT
ecosystems, enabling automation, real-time alerts, and water budgeting features. The system
can be applied in multi-unit apartment complexes, commercial establishments, and industrial
facilities where monitoring and predicting appliance-level water consumption is critical for
20 resource optimization.
[0016] These and other aspects of the embodiments herein will be better appreciated
and understood when considered in conjunction with the following description and the
accompanying drawings. It should be understood, however, that the following descriptions,
while indicating preferred embodiments and numerous specific details thereof, are given by
25 way of illustration and not of limitation. Many changes and modifications may be made
within the scope of the embodiments herein without departing from the spirit thereof, and the
embodiments herein include all such modifications.
BRIEF DESCRIPTION OF THE DRAWINGS
7
[0017] The embodiments herein will be better understood from the following detailed
description with reference to the drawings, in which:
[0018] FIG. 1 illustrates a system for appliance-level disaggregation and prediction of
water consumption in real-time according to some embodiments herein;
[0019] FIG. 2 illustrates an exploded view of aremote server of the 5 system of FIG.1
according to some embodiments herein;
[0020] FIG. 3 illustrates an exemplary deployment of the system of FIG.1 into a
premises according to some embodiments herein;
[0021] FIG. 4is a process flow diagram that illustrates training of a machine learning
10 model of the system of FIG.1 according to some embodiments herein;
[0022] FIGS. 5A-B are graphs illustrating performance of the system for appliancelevel
disaggregation and prediction of water consumption in real-timeaccording to some
embodiments herein.
[0023] FIG. 6 illustrates a confusion matrix that underscores the high accuracy of the
15 LSTM model in distinguishing between filling, usage, and no activity phases according to
some embodiments herein;
[0024] FIGS. 7A-B are graphs illustrating the comparison between the original water
level data and the estimations generated by the Long Short-Term Memory (LSTM) model
according to some embodiments herein;
20 [0025] FIGS. 8A-B are flow diagrams that illustrate a method forappliance-level
disaggregation and prediction of water consumption in real-timeaccording to some
embodiments herein; and
[0026] FIG. 9 is a schematic diagram of a computer architecture in accordance with
the embodiments herein.
25 DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS
[0027] The embodiments herein and the various features and advantageous details
thereof are explained more fully with reference to the non-limiting embodiments that are
illustrated in the accompanying drawings and detailed in the following
8
description.Descriptions of well-known components and processing techniques are omitted
so as to not unnecessarily obscure the embodiments herein. The examples used herein are
intended merely to facilitate an understanding of ways in which the embodiments herein may
be practiced and to further enable those of skill in the art to practice the embodiments
herein.Accordingly, the examples should not be construed as limiting 5 the scope of the
embodiments herein.
[0028] As mentioned, there remains a needto provide an IoT and Machine Learningbased
water load disaggregation system designed to monitor and predict water usage at the
appliance level in a premises. By utilizing a network of ultrasonic sensors and a central water
10 level monitoring device, the system collects real-time data on water consumption. This data
is processed using machine learning algorithms, which disaggregate the total water usage
into individual appliance consumption. Referring now to the drawings, and more particularly
to FIGS. 1 through 9, where similar reference characters denote corresponding features
consistently throughout the figures, there are shown preferred embodiments.
15 [0029] FIG. 1 illustrates a system for appliance-level disaggregation and prediction of
water consumption in real-timeaccording to some embodiments herein.The system 100
includes one or more ultrasonic sensors 102A-N, a smart water level monitoring device 104,
a processor 106, a remote server 108, and a network 110.The one or moreultrasonic sensors
102A-N is operatively mounted at pipelines of individual water-using appliances/devices like
20 flushes, washing machines, jet sprays, geysers, etc. in a premises(e.g., a house). The one or
more ultrasonic sensors 102A-N is configured to measure water volume or flow rate as water
passes through the pipelinesin real-time. The smart water level monitoring device 104 is
positioned in a water tank and configured to continuously measure/monitor the water level in
the watersupplytank. The processor 106 is communicatively coupled to the one or more
25 ultrasonic flow sensors 102A-Nand the smart water level monitoring device 104 via a
network 110. The network 110 is a wired network or a wireless network. The processor 106
may be a part of an IoT node. The processor 106 is configured to receive real-time data from
the one or moreultrasonic sensors 102A-N and the smart water level monitoring device 104.
9
Thereal-time data includes flow rates from individual water-using appliances and water level
changes in the water supply tank.The processor 106 is configured to collect and process the
real-time data from the one or moreultrasonic sensors 102A-N and the smart water level
monitoring device 104 and communicate to the remote server 108 through the network 110.
The remote server 108 may be a cloud server.The one or more ultrasonic 5 sensors 102A-N
outputs a pulse signal that is proportional to the flow rate of the water passing through it. The
processor 106 reads the pulse signals generated by the one or more ultrasonic sensors 102AN
and determines the corresponding flow rate.
[0030] The remote server 108 implementsa machine learning model 112, pre-trained
10 on a processed dataset of labeled water usage events paired with corresponding flow data and
water level changes and analyzes the incoming data stream. The processed dataset is obtained
by one or more ultrasonic sensors 102A-N. The one or more ultrasonic sensors 102A-N is
flow sensors. The flow sensors may be a Digital Water Flow Sensor (e.g., YF-S201). The
flow sensorsprovide the ground truth data or processed dataset needed for training the
15 machine learning model 112. This datasetis labeled and used to train the machine learning
model 112 to recognize patterns and predict flow-related outcomes accurately. The data
gathered by the flow sensors is often used to extract features that are used for training the
machine learning model 112. These features may include flow rate, velocity, pressure, and
other dynamics. During the training, the flow sensors are used to calibrate the machine
20 learning model 112 by providing a reliable baseline of how flow behaves under different
conditions. Once the machine learning model 112 is trained, the system 100 can predict flow
behavior without needing real-time sensor input. The flow sensors assist in validating the
accuracy of the model by comparing the predictions of the machine learning model 112
against the actual sensor readings. This validation process ensures that the machine learning
25 model 112 can generalize well to unseen data. Over time, the machine learning model 112 is
continuously updated with new data, allowing it to learn and improve its accuracy. The
machine learning model 112 identifies the specific appliance responsible for the current
water usage based on the analysis of flow patterns and water level changes that facilitate
10
disaggregating the total water consumption into individual appliance consumption, providing
detailed insights into water usage distribution. The machine learning model 112 predicts
future water consumption for each appliance based on historical usage patterns and real-time
water level changes, empowering users to make informed decisions about water conservation
efforts. The remote server 108 is configured to communicate the detailed insights 5 into water
usage distribution and the predicted future water consumption for each appliance to the
user114 of the premises through a user device associated with the user 114.
[0031] In some embodiments, the one or more ultrasonic sensors 102A-N and the
smart water level monitoring device 104 are synchronized to capture data points every 2
10 seconds. The collected data includes timestamps, water level readings, flow rates, and
appliance activity labels. This approach provides accurate and detailed monitoring of water
usage patterns.The continuous data collection is crucial for subsequent data analysis and
estimating appliance water loads.
[0032] In some embodiments, the processor 106 is an IoT edge node configured to
15 buffer and compress real-time sensor data prior to transmitting to the remote server108. In
some embodiments, the remote server 108 is configured to generate anomaly alerts based on
deviations between predicted and actual usage, indicating potential leaks or inefficiencies.
[0033] FIG. 2 illustrates an exploded view of aremote server 108 of the system 100 of
FIG.1 according to some embodiments herein. The remote server 108 includes a database
20 202, a real-time data receiving module 204, a data processing module 206, a total
waterusagedisaggregation module 208, a water-using appliance identification module 210, a
future water usage prediction module 212, a data transmission module 214 and the machine
learning model 112. The real-time data receiving module 204 is configured to receive realtime
data from the one or moreultrasonic sensors 102A-N and the smart water level
25 monitoring device 104. The data processing module 206 is configured to process the realtime
data from the one or moreultrasonic sensors 102A-N and the smart water level
monitoring device 104.The total water usage disaggregation module 208disaggregates,using
the machine learning model112, a total water usage into individual water-using appliance
11
usage by analyzing variation in the flow rate and water level changes patterns of water
passing through the pipelines in real-time. The water-using appliance identification module
210identifies in real-time the specific water-using appliance responsible for current water
usage by applying a trained appliance classification model to the received flow and water
level data. The future water usage prediction module 212predicts using a 5 multivariate time
series forecasting model comprising a Long Short-Term Memory (LSTM) neural network, a
future water usage for each water-using appliance based on historical usage patterns and
ongoing tank water level changes.The data transmission module 214 transmits the
disaggregated appliance-level usage data and the predicted future usage profile to a user
10 devicefor enabling water conservation actions by a user 114.
[0034] In some embodiments,the machine learning model 112 is trained using
historical data including a processed dataset of labeled water usage events. The dataset
includes flow rate, velocity, pressure, and water level changes obtained from flow sensors.
The flow sensors provides ground truth data used for calibrating and validating the model to
15 accurately recognize flow patterns, disaggregate appliance-level consumption, and predict
future usage behaviors. The machine learning model 112 includes at least one of a Long
Short-Term Memory (LSTM) neural network or a one-dimensional convolutional neural
network (Conv1D) configured to model temporal patterns in multivariate water usage data.
[0035] FIG. 3 illustrates an exemplary deployment of the system of FIG.1 into a
20 premises according to some embodiments herein. The system 100 includes one or more
ultrasonic flow sensors 102A–J installed at pipelines of individual water-using appliances
across the premises. As shown in FIG. 3, the flow sensors 102A–Jare mounted at the kitchen
sink, RO water purifier, bathroom cabinet, jet spray, flush, geyser, washing machine, and tap
outlets. A smart water level monitoring device 104 is positioned in the overhead water tank
25 located on the terrace to continuously measure the storage level.The flow sensors 102A–Jare
configured to measure real-time water volume or flow rate through respective pipelines,
while the smart water level monitoring device 104 tracks water availability in the water
supply tank. The collected data is transmitted via a network 110 to a processor 106, which
12
may be implemented as an IoT edge node. The processor 106 aggregates flow and level data
and communicates it to a remote server 108.The remote server 108executes a machine
learning model 112 trained on labeled flow datasets to recognize appliance-specific usage
patterns. Based on analysis of incoming flow signatures and water level changes, the
machine learning model 112 disaggregates total household consumption into 5 appliance-level
insights. The system thereby enables detailed monitoring of tank refilling, water storage
status, and per-appliance consumption across kitchen, toilet, and bathroom zones, providing
users 114 with actionable information for water conservation and anomaly detection.
[0036] FIG. 4is a process flow diagram that illustrates training of a machine learning
10 model 112 of the system 100 of FIG.1 according to some embodiments herein.The system
100 begins with a learning-based water load monitoring, where data is collected from two
types of sources:(i) smart water level monitoring devicecontinuously measures the water
level in the water supply tank, and (ii) one or moreultrasonic flow sensors 102A-N installed
on pipelines of respective water-using appliances such as flushes, washing machines,
15 geysers, and jet sprays, and record the water flow rate each time an appliance is used.The
smart water level monitoring device captures water level readings from the water supply
tank.The one or moreultrasonic flow sensors 102A-N capture real-time flow rates and
timestamped activity of each appliance.Both sources are synchronized to record data every 2
seconds, ensuring accuracy and consistency.The raw dataset includes timestamps, water
20 levels, flow rates, and appliance activity labels.The collected data undergoes preprocessing to
remove noise, correct errors, and align readings from both the tank and appliance-level flow
sensors.After preprocessing, the cleaned dataset contains reliable water usage patterns, ready
for model training.Using the activity logs from flow sensors, each dataset entry is labeled
according to the specific appliance responsible for the water usage event.These appliance25
level labels are essential for training the machine learning model 112 to recognize flow
patterns and disaggregate total water usage into appliance-specific usage.
[0037] The cleaned and labeled dataset is fed into a machine learning model112
implemented on the remote server 108.The machine learning model 112 can include Long
13
Short-Term Memory (LSTM) or 1D Convolutional Neural Networks (Conv1D), which are
specifically designed to learn temporal and sequential patterns in multivariate data.The
training enables the machine learning model 112 to recognize flow signatures of appliances,
disaggregate water usage, and learn temporal behavior for prediction.Once trained, the
machine learning model 112 is tested with unseen water usage data 5 to validate its
accuracy.Testing ensures that the model can correctly identify appliances and predict water
consumption patterns under real-world conditions.The outputs of the machine learning model
112 are transformed into meaningful insights for the user.This includes appliance-level
disaggregation (which appliance is using how much water) and future consumption
10 predictions based on historical patterns and current water levels.The results are transmitted to
a user device, enabling the user to monitor real-time consumption and take conservation
measures.
[0038] The system provides a complete appliance-level breakdown of water usage
and predictive consumption profiles, thereby supporting efficient water management, leakage
15 detection, and conservation actions.
[0039] The below tables 1 and 2 illustrate the number of data points collected from
various water-related activities, both before and after resampling. Resampling is a data
processing technique employed to reduce and balance datasets, typically for the purpose of
improving the performance of machine learning models. Before resampling, categories such
20 as "Water Level" and "No Activity" contained disproportionately large numbers of data
points (159,016 and 120,598, respectively). This imbalance may lead to biased analysis or
training outcomes. After resampling, these numbers are significantly reduced, with "Water
Level" dropping to 33,583 and "No Activity" to 21,145. The total number of data points
across all appliances is also reduced from 8,798 to 2,898. This balanced distribution after
25 resampling helps to prevent any single category from dominating the dataset, leading to more
accurate and fair model predictions.
[0040] Table 1: Number of datapoints before resampling.
Name Data points
14
Water level 159016
Flush 2204
Washing machine 2816
Geyser 1810
Jet spray 1968
Filling 29628
No Activity 120598
Total of appliances 8798
[0041] Table 2: Number of data pointsafter resampling
Name Data points
Water Level 33583
Flush 726
Washing Machine 1002
Geyser 597
Jet Spray 573
Filling 9540
No Activity 21145
Total of appliances 2898
[0042] CSV files of the real-time data areloaded to pandas DataFrames by (i)
removing unnecessary columns from data such as entry_id, longitude, latitude, etc, (ii)
renaming columns ex: - filed1 to flow_rate, filed2 to volume_per_usage,(iii) setting the
timestamp as an index and converting the timestamp to standard DateTime 5 format. The
rolling mean (Moving Average Filter) of window = 10 is used to remove the noise and
smoothen the data. For removing outliers, z-score thresholding with a threshold of less than 3
is used. A moving average filter is used to smooth out short-term fluctuations and highlight
longer-term trends or cycles in the data. This is achieved by calculating the average of a fixed
10 number of consecutive data points (window size) and using this average as the value for the
15
central data point in the window. The window then slides over to the next set of data points,
repeating the process.
[0043] Example: Given a time series: [3,5,8,10,12,14,18,21,23,25] and a window size
of 3, the moving average at the third position would be calculated as: Moving Average = 3 +
5 + 8 / 3 = 5.33. This process is repeated for each point in the series, 5 producing a smooth
series.
[0044] The outliers are removed with Z-Score thresholding. Z-score thresholding is a
statistical method to identify and remove outliers. The z-score represents the number of
standard deviations a data point is from the mean of the series. Data points with z-scores
10 above a certain threshold are considered outliers and are removed or replaced. Z-Score = -
µ / s Example: For a time series with a mean of 10 and a standard deviation of 2, a data point
with a value of 16 may have a z-score of: Z-Score = 16 – 10 / 2 = 3. If the threshold is set to
2, this point maybe considered an outlier and can be removed or replaced.
[0045] The data is resampled to a specified frequency, to align the data based on
15 timestamps. Resampling involves changing the frequency of time series data, either by
aggregating data to a lower frequency (downsampling) or interpolating data to a no activity
frequency (up-sampling). This helps in standardizing the time series data for consistent
analysis. The steps include the following: (1) Define Frequency: Specify the desired
frequency for resampling,(2) Apply Resampling: Use a resampling method appropriate for
20 the data (e.g., mean, sum, interpolation) to convert the data to the desired frequency.
Example: Given water level data recorded every 2 seconds, resampling to 10-second
intervals can be done using the mean value within each 10-second window.
[0046] The appliancedata are assigned with theirrespective labels for each data point
and merged to a single DataFrame based on the Timestamp of water level data and merged
25 appliances data. Respective labels are assigned to the water level data. When there is no label
at the Timestamp, a ‘no_activity’ label is assigned to the water level data. Example: if at
Timestamp 6 June 2024 11:55:30 == ‘flush’ label in merge appliances data. At Timestamp 6
June 2024 11:55:30 in water level data the label will be a ‘flush’ label. Outliers in the labels
16
can occur due to various reasons, such as sensor noise, network issues, or data loss in
appliancedata. The outliers are corrected by comparing each label with the majority of labels
within a moving window. The steps include the following: (i) Define a Moving Window:
Specify the size of the moving window (e.g., 5 consecutive data points). (ii) Identify the most
common label: For each window, determine the most common label 5 (mode) within the
window. (iii) compare Labels: Compare each label within the window to the most common
label.
[0047] Table 3: uncorrected labels
Index Label
0 Flush
1 Flush
2 No Activity
3 Flush
4 Flush
5 Flush
6 Flush
[0048] Using a window of size , Window 0-4: Labels: [Flush, Flush, No Activity,
10 Flush, Flush]. The most common label: Flush. Corrections: The label 'No Activity' at index 2
is corrected to 'Flush'.
[0049] Window 1-5: Labels: [Flush, No Activity, Flush, Flush, Flush]. The most
common label: Flush. Corrections: The label 'Flush' at index 3 remains corrected to 'Flush'.
[0050] Table 4: corrected labels
Index Label
0 Flush
1 Flush
2 Flush
3 Flush
4 Flush
17
5 Flush
6 Flush
[0051] The processed datasets are saved to the CSV file.The datasets are preprocessed
for training the machine learning model 112. The pre-processing of the datasets for
training the machine learning model 112 includes creating Time Series Windows, Label
Encoding and Standard Scaling.
[0052] Creating Time Series Windows: In time series analysis, creating 5 windows of
data is essential to capture temporal dependencies and prepare the data for various machine
learning models. This technique involves segmenting the time series data into smaller,
overlapping sequences, or "windows," allowing the machine learning models to learn from
patterns within these smaller segments.
10 [0053] Definition and Process: Let X be the time series data with n samples, where
each sample has m features: X = {1, 2, … , }. Assume the window size is w. Each
window is a segment of w consecutive samples: = {, +1, … , +-1}.This process is
repeated by sliding the window with a step size s (often s = 1 for overlapping windows):
{1,2, … , -+1 .For example:For time series data = 1,2,3,4,5 with labels =
15 0,0,1,1,0: 1 = {1,2,3} ? label = 1, 2 = {2,3,4} ? label = 1, 3 = {3,4,5} ? label = 0. So,
the labeled windows are: {( W1, 1), (W2, 1), (W3, 0)} = { ({1, 2, 3}, 1), ({2, 3, 4}, 1),
({3, 4, 5}, 0) }.
[0054] Label Encoding: Label encoding is the process of converting categorical
labels into numerical values. An integer value is assigned to each unique label. For example:
20 labels = [ 'flush', 'geyser', 'washing_machine'], 'flush' ? 0, 'geyser' ? 1, 'washing_machine'
? 2.
[0055] Standard Scaling: Feature scaling is essential to ensure that all input features
contribute equally to the result and to accelerate the convergence of the learning algorithm.
Standard scaling transforms the features so that they have a mean of zero and a standard
25 deviation of one. = -
/ . Where: is the original feature value, µ is the mean of the
feature, s is the standard deviation of the feature, is the scaled feature value. For example:
18
Features = [10,20,30,40,50], µ = 10 + 20 + 30 + 40 + 50 / 5 = 30, s = v (10 - 30) 2 + (20 -
30) 2 + (30 - 30) 2 + (40 - 30) 2 + (50 - 30) 2 / 5 = v200 ˜ 14.14. 1 = 10 - 30 /14.14 ˜
-1.41, 2 = 20 - 30 /14.14 ˜ -0.71, 3 = 30 – 30 /14.14 = 0, 4 = 40 – 30 /14.14 ˜ 0.71, 5 =
50 - 30 /14.14 ˜ 1.41 and scaled features = [-1.41, -0.71,0,0.71,1.41].
[0056] The system addresses a Multivariate Time Series Classification 5 problem,
where it is crucial to consider past temporal features of data points to accurately predict
current labels. The water level patterns exhibit a time-dependent behavior with a consistent
upward trend as each new data point is recorded. To effectively capture these evolving trends
and intricate patterns, the system utilizes sequential models such as Long Short-Term
10 Memory (LSTM) and 1D Convolutional Neural Networks (Conv1D). These models are
specifically designed to interpret the sequential nature of the data, enabling them to identify
complex relationships between historical patterns and real-time data interactions.
[0057] Long Short-Term Memory (LSTM) is a type of artificial recurrent neural
network (RNN) architecture used in the field of deep learning. Unlike standard feedforward
15 neural networks, LSTMs have feedback connections, allowing them to exploit temporal
dependencies across sequences of data. LSTM is designed to handle the issue of vanishing or
exploding gradients, which can occur when training traditional RNNs on sequences of data.
LSTM networks introduce memory cells, which have the ability to retain information over
long sequences. Each memory cell has three main components: an input gate, a forget gate,
20 and an output gate. These gates help regulate the flow of information in and out of the
memory cell.
[0058] Gates of LSTM:(i) Forget Gate: It determines what information to discard
from the cell state and it takes input (current time step and previous hidden state) and
produces a number between 0 and 1 for each number in the cell state, where 1 represents
25 “completely keep this” while 0 represents “completely get rid of this”,(ii) Input Gate: it
decides what new information to store in the cell state and it consists of two parts,(1) A
sigmoid layer (the “input gate layer”) decides which values to update and(2) A tanh layer
(which creates a vector of new candidate values to add to the cell state),and (iii) Output Gate:
19
itdetermines the next hidden state based on the updated cell state and filters the information
that the LSTM may output based on the updated cell state.
[0059] The LSTM includes a cell state,a hidden state, dropout layers and a dense
output layer. The cell stateruns straight down the entire chain of the LSTM, with only some
minor linear interactions. It’s the core differentiator in LSTMs that allows 5 them to maintain
and control long-term dependencies. The hidden state is the LSTM’s output at a particular
time step based on the cell state.
= s(· [h
-1,
] + ),
= s(· [h
-1,
] + ),
~ =
tanh(· [h
-1,
] + ),
=
*
-1 +
*
~,
= s(· [h
-1,
] + ), h
=
*tanh(
).
Where: ?
is the input at time step t, ? h
is the hidden state, ?
is the cell state, ?
,
,
10
are the forget, input, and output gates, respectively, ? denotes the sigmoid function, and
\tanh denotes the hyperbolic tangent function. The dropout layers randomly set a fraction of
input units to 0 at each update during training to prevent overfitting. If p is the dropout rate,
the dropout operation can be expressed as, output = input · mask, mask ~ Bernoulli(1 - ).
The dense output layer is a fully connected layer with a softmax activation function for
15 multi-class classification: softmax( ) = / S , where z is the input to the softmax
function for the class .
[0060] The Conv1D operation involves a filter (or kernel) that slides over the input
sequence, performing a convolution operation to extract features from the data. It is used
primarily for analyzing sequential data where the data points are ordered in one dimension.
20 By applying filters over the temporal dimension, the network can learn to recognize patterns
and dependencies within the sequence.
[0061] The input to a Conv1D layer is typically a 2D array with the shape (timesteps,
features), where: timesteps refer to the length of the input sequence, and features refer to the
number of features at each timestep.A filter is a smaller array with a specific width (kernel
25 size) that slides over the input sequence. Each filter learns to detect different patterns in the
sequence. The convolution operation can be described mathematically. Let: ? x be the input
sequence. ? w be the filter weights. ? b be the bias term. ? K be the kernel size. ? y[t] be
the output at position t. The output is calculated as:
[0062] Where, kernel size
much the filter moves at each step. A stride of 1 means the filter moves one step at a time,
while a stride of 2 means it moves two steps at a time.
dimensions of 5 the output. The v
the size of the output. Same
is the same as the input size.Example of
is N and the kernel size is K, t
ensures the output length is the same as the input length N.
10 [0063] Activation Function (Relu): After the convolution operation, the output is
typically passed through an activation function like ReLU (Re
= max(0, ).
[0064] Pooling Layers: Pooling layers are often used after convolutional layers to
down sample the output, reducing its dimensionality and computational load while
15 preserving important features.
[0065] MaxPooling1D: Takes the maximum value from each segment of the input
sequence. 1(
[0066] Example of Pooling
pooling operation will yield [3,4].
20 [0067] Flatten Layer: Converts the 2D output of the convolutional and pooling layers
into a 1D vector. If the input to the Flatten layer has a shape (timesteps, features)
transformed into a single vector of size timesteps × features.
[0068] Dense Layers: Fully connected layers combine the features learned by the
convolutional layers. Dense(
25 W is the weight matrix, x is the input vector, and b is the bias term.
20
sizedetermines the width of the filter. The stride
The padding controls the spatial
valid padding means no padding is added, which can reduce
padding is the padding which is added to ensure the output size
padding: valid padding: if the input sequence length
the output length will be N-K+1. Same padding:
Rectified Linear Unit)
) = max([: + ]).
Pooling: For a pool size of 2: If the input is [3,1,4,2] the max
) = s( + ) Here, s is the activation function (e.g., ReLU),
tridedetermines how
ontrols o adding: the padding
ctified Unit). ()
it is
[0069] Dropout Layers: Dropout layers randomly set a fraction of input units to 0 at
each update during training to prevent overfitting. If p is the dropout rate, the dropout
operation can be expressed as.
[0070] output = input
[0071] Output Layer: A fully connected layer with a softmax activation 5 function for
multi-class classification:
[0072] The model training
model.
10 [0073] Compiling the Model: loss function: categorical cross
classification)
[0074] Where: ? N is the number of samples,
the binary indicator (0 or 1) if
15 ?!^,"is the predicted probability of the sample
[0075] Adam optimizer updates the parameters based on the estimates of first and
second moments of the gradients:
21
· mask, mask ~ Bernoulli(1 – )
, where z is the input to the softmax function for the
includes compiling, defining call backs and training the
cross-entropy (for multi
? C is the number of classes,
class label c is the correct classification for the sample
? being in class c.
here ?.
multi-class
?!,"is
?,
[0076] where: ?
is the gradient at time step t,
of the gradient and its square,
is a small constant for numerical stability.
[0077] Defining Callbacks: (i) ModelCheckpoint:
training based 5 on validation loss
the validation loss plateaus
improvement in validation loss for a specified numbe
[0078] The models are trained using the training data, with a portion set aside for
validation. The training involves multiple epochs, where the model learns to minimize the
10 loss function and improve accuracy.
[0079] Model Evaluation includes
trained, the models are used to
that the model has not seen during training, providing an unbiased assessment of its
performance. (2) Accuracy: Accuracy meas
15 instances out of the total instances.
[0080] Where:

be the test data
predictions made by the model
function, which is 1 if the condition inside is true and 0 otherwise.
20 [0081] (3) Precision: Precision for class c is the ratio of true positive predictions to
the total predictions made for class c :
[0082] (4) Recall (Sensitivity): recall for class c
predictions to the total actual instances of class c:
22
?#
and $
are the moving averages
? ß1 and ß2 are hyperparameters, ? a is the learning rate,
saves the best model during
loss, (ii) ReduceLROnPlateau: reduces the learning rate when
plateaus, and (iii) EarlyStopping: Stops training when there is no
number of epochs.
(1) Predicting on Test Data: Once the models are
make predictions on a separate test set. This set contains data
measures the proportion of correctly predicted
data, !true be the true labels for the test data.
model, N is the total number of test instances, %(·
is the ratio of true positive
? ?
aves educes ures data.,!pred be the
·) is the indicator
[0083] (5) F1 Score: The F1 score for class c is the harmonic mean of precision and
recall:
5
[0084] FIGS. 5A-B
appliance-level disaggregation and prediction of water consumption
some embodiments herein. The graph shown in FIG.
measurements over time, highlighting periods of filling, usage, and no activity
shows the estimations generated by the LSTM model, which closely mirror 10 the original data.
The comparison of these graphs demonstrates the
level activities.
[0085] FIG. 6 illustrates a
the LSTM model in distinguishing between filling, usage, and no activity phases
15 some embodiments herein. The confusion matrix visually represents the performance of the
classification model in distinguishing between the different water level activities: filling, no
activity, and usage. The diagonal values indicate correct predictions, with the model
accurately identifying 1,878 instances of filling, 2,220 instances of no activity, and 265
instances of usage. Off-diagonal values show misclassifications, such as 51 cases where
20 filling was incorrectly predicted as no activity, and 116 cases where usage was misclassified
as no activity. The matrix highlights the model's strong performance, par
predicting "no activity" and "filling" states, with fewer misclassifications compared to other
categories.
[0086] FIGS. 7A-B are graphs illustrating the comparison between the original water
25 level data and the estimations generated by the Long Short
23
are graphs illustrating the performance of the system
in real-
. 5A depicts actual water level
system's ability to accurately predict water
confusion matrix, which underscores the high accuracy of
del Short-Term Memory (LSTM)
100for
-timeaccording to
A activity. FIG. 5B
's according to
particularly in
24
modelaccording to some embodiments herein. Both graphs 7A and 7B show the water level
in centimeters over time, with labels indicating different activities such as filling, no activity,
jet spray, geyser, flush, and washing machine usage. The LSTM estimation graph 7B, closely
follows the original data, demonstrating the model's effectiveness in predicting water level
trends and corresponding activities. Despite some minor discrepancies, 5 the LSTM model
captures the overall patterns and transitions between activities, indicating its strong
performance in estimating the water levels based on historical data.The results, as
demonstrated in the graphs and confusion matrix, highlight the reliability of the smart water
level monitoring device 104 (i.e.) ultrasonic water level sensor in accurately estimating
10 appliance water usage. Utilizing the Long Short-Term Memory (LSTM) deep learning
model, the analysis is able to successfully identify specific appliances consuming water, such
as the washing machine, geyser, and jet spray, and measure their individual water
consumption. By analyzing patterns in the water level data, the model distinguished between
different appliances and their specific usage behaviors. Achieving an impressive accuracy
15 rate of 88%, the LSTM model has proven highly effective in both monitoring and estimating
water usage for each appliance. This high level of accuracy not only ensures dependable
performance but also underscores the model's potential for optimizing water management
and detecting anomalies in appliance usage, which could lead to significant water
conservation and efficiency improvements.
20 [0087] FIGS. 8A-B are flow diagrams that illustrate a method for appliance-level
disaggregation and prediction of water consumption in real-time according to some
embodiments herein.At step 802,real-time flow rate data is received from one or more
ultrasonic flow sensors 102A-N mounted on pipelines of respective water-using appliances
within a premises using a processor 106. At step 804,real-time water level data is received
25 from a smart water level monitoring device 104 positioned in a water supply tank using the
processor 106.At step 806,the real-time flow rate data and water level data is transmittedto a
remote server 108using the processor 106 via a network 110. The remote server 108
implements a machine learning model 112 trained using historical labeled flow data.At step
25
808,a total water usage is disaggregated into individual water-using appliances usage using
the processor 106 and the machine learning model 112 by analyzing variation in the flow rate
and water level changes patterns of water passing through the pipelines in real-time.At step
810, the specific water-using appliance responsible for current water usage is identified in
real-timeby applying a trained appliance classification model to the received 5 flow and water
level data.At step 812, a future water usage for each water-using appliance is predicted, using
a multivariate time series forecasting model comprising a Long Short-Term Memory
(LSTM) neural network, based on historical usage patterns and ongoing tank water level
changes.At step 814, the disaggregated appliance-level usage data and the predicted future
10 usage profile is transmittedto a user device for enabling water conservation actions by a user.
[0088] In some embodiments, predicting future water consumption for each waterusing
appliance includes generating a time-stamped forecast of water usage for each
appliance for a predefined future time interval.In some embodiments, disaggregated water
usage for each appliance is classified into activity categories comprising at least one of:
15 filling, usage, and no activity, and the machine learning model generates confidence scores
associated with each classification.In some embodiments, the machine learning model is
periodically updated with newly labeled flow rate and water level data to continuously
improve the accuracy of appliance identification and usage prediction.
[0089] In some embodiments, the method includes displaying, on a mobile
20 application associated with the user device, a real-time dashboard presenting individual
appliance-level water usage data and corresponding predictive insights generated by the
machine learning model 112. In some embodiments,the user device includes at least one of a
mobile phone, smart speaker, or connected dashboard configured to receive push
notifications based on the machine learning model's output.
25 [0090] The method enables efficient water usage monitoring and management by
leveraging IoT nodes equipped with ultrasonic sensors and the computational power of
remote servers hosting machine learning models. The method provides a comprehensive
solution for monitoring and optimizing water consumption in various environments.Unlike
26
manual monitoring methods, the method provides real-time data on water consumption at a
granular level, specifically identifying usage at individual appliances. This enhances
accuracy and eliminates the need for manual meter readings, which can be time-consuming
and prone to human error. Compared to smart water meters that often require expensive
replacements and plumbing modifications, themethod adapts to 5 existing plumbing
infrastructure. It offers detailed consumption data without the high costs associated with
smart meter installations. The method goes beyond static water consumption labels by
providing real-time measurements that reflect actual usage patterns. By identifying abnormal
flow patterns, the method can detect leaks or inefficiencies, allowing for timely interventions
10 and optimization of water usage. The use of a machine learning model 112 enables
continuous learning and improvement of water usage predictions, providing insights that
allow users to take proactive measures in water conservation. The method offers detailed
insights into water consumption and predictions of future usage, allowing homeowners to
monitor water usage patterns easily through their devices. This leads to informed decisions
15 on water conservation and efficiency efforts.
[0091] A representative hardware environment for practicing the embodiments herein
is depicted in FIG. 9 with reference to FIGS. 1 through 8. This schematic drawing illustrates
a hardware configuration of a server 118 /computer system in accordance with the
embodiments herein. The server 118 /computer includes at least one processing device 10
20 and a cryptographic processor 11. The special-purpose CPU 10 and the cryptographic
processor (CP) 11 may be interconnected via system bus 14 to various devices such as a
random access memory (RAM) 15, read-only memory (ROM) 16, and an input/output (I/O)
adapter 17. The I/O adapter 17 can connect to peripheral devices, such as disk units 12 and
tape drives 13, or other program storage devices that are readable by the system. The server
25 118 / computer can read the inventive instructions on the program storage devices and follow
these instructions to execute the methodology of the embodiments herein. The server
118/computer system further includes a user interface adapter 20 that connects a keyboard
18, mouse 19, speaker 25, microphone 23, and/or other user interface devices such as a touch
27
screen device (not shown) to the bus 14 to gather user input. Additionally, a communication
adapter 21 connects the bus 14 to a data processing network 26, and a display adapter 22
connects the bus 14 to a display device 24, which provides a graphical user interface (GUI)
30 of the output data in accordance with the embodiments herein, or which may be embodied
as an output device such as a monitor, printer, or transmitter, for example. 5 Further, a
transceiver 27, a signal comparator 28, and a signal converter 29 may be connected with the
bus 14 for processing, transmission, receipt, comparison, and conversion of electric or
electronic signals.
[0092] The foregoing description of the specific embodiments will so fully reveal the
10 general nature of the embodiments herein that others can, by applying current knowledge,
readily modify and/or adapt for various applications such specific embodiments without
departing from the generic concept, and, therefore, such adaptations and modifications
should and are intended to be comprehended within the meaning and range of equivalents of
the disclosed embodiments.It is to be understood that the phraseology or terminology
15 employed herein is for the purpose of description and not of limitation. Therefore, while the
embodiments herein have been described in terms of preferred embodiments, those skilled in
the art will recognize that the embodiments herein can be practiced with modification within
the spirit and scope. ,CLAIMS:I/We claim:
1. A system (100) forappliance-level disaggregation and prediction of water consumption in
real-time, the system (100) comprising:
a plurality of ultrasonic flow sensors (102A-N), that is operatively mounted on pipelines
of respective water-using appliances within a premisesand is configured to measure flow
rates of water passing through the pipelines 5 in real-time;
a smart water level monitoring device (104) that ispositioned in a water supply tank and
configured to continuously measure water levelin the water supply tank;
a processor (106) that is communicatively coupled to the plurality of ultrasonic flow
sensors (102A-N) and the smart water level monitoring device (104) via a network (110),
10 wherein the processor (106) is configured to preprocess and transmit real-time sensor data
comprising flow rates from individual water-using appliances and water level changes in the
water supply tank; and
a remote server (108) that is communicatively coupled to the processor (106), the remote
server (108) comprising a memory;
15 characterized in that, wherein the remote server (108) implementsa machine learning model
(112) trained using historical labeled flow data and is configured to:
receive, usingthe processor (106), time-series flow rate data from the pipelines and
water level data from thewater supply tank;
disaggregate, using the processor (106) and the machine learning model(112), a total
20 water usage into individual water-using appliance usage by analyzingvariation in theflow rate
and water level changespatterns ofwater passing through the pipelines in real-time;
identify in real-time the specific water-using appliance responsible for current water
usageby applying a trained appliance classification model to the received flow and water
level data;
25 predict, using a multivariate time series forecasting model comprising a Long Short-
Term Memory (LSTM) neural network, a future water usage for each water-using appliance
based on historical usage patterns and ongoing tank water level changes; and
transmit, to a user device, the disaggregated appliance-level usage data and the
predicted future usage profile for enabling water conservation actions by a user (114).
29
2. The system (100) as claimed in claim 1, wherein the machine learning model (112) is
trained using historical data comprising a processed dataset of labeled water usage events,
wherein the dataset comprisesflow rate, velocity, pressure, and water level changes obtained
from flow sensors, wherein the flow sensors provides ground truth data used 5 for calibrating
and validating the model to accurately recognize flow patterns, disaggregate appliance-level
consumption, and predict future usage behaviors, and wherein the machine learning model
(112) comprises at least one of a Long Short-Term Memory (LSTM) neural network or a
one-dimensional convolutional neural network (Conv1D) configured to model temporal
10 patterns in multivariate water usage data.
3. The system (100) as claimed inclaim 1, wherein the plurality of ultrasonic flow sensors
(102A-N) are configured to generate pulse signals proportional to the flow rate of water, and
the processor (106) converts the pulse signals into digital flow rate data.
15
4. The system (100) as claimed inclaim 1, wherein the smart water level monitoring device
(104) comprises an ultrasonic sensor configured to detect minute changes in water level
every two seconds.
20 5. The system (100) as claimed inclaim 1, wherein the processor (106) is an IoT edge node
configured to buffer and compress real-time sensor data prior to transmitting to the remote
server(108).
6. The system (100) as claimed inclaim 1, wherein the remote server (108) is configured to
25 generate anomaly alerts based on deviations between predicted and actual usage, indicating
potential leaks or inefficiencies.
7. A computer-implemented method for appliance-level disaggregation and prediction of
water consumption in real-time, the method comprising:
30
receiving, using a processor (106), real-time flow rate data from a plurality of
ultrasonic flow sensors (102A-N) mounted on pipelines of respective water-using appliances
within a premises;
receiving, using the processor (106), real-time water level data from a smart water
level monitoring device (104) positioned in a 5 water supply tank;
transmitting, using the processor (106) via a network (110), the real-time flow rate
data and water level data to a remote server (108) comprising a memory, wherein the remote
server (108) implements a machine learning model (112) trained using historical labeled flow
data;
10 disaggregating,using the processor (106) and the machine learning model, atotal
water usage into individual water-using appliances usage by analyzingvariation in the flow
rate and water level changespatterns ofwater passing through the pipelines in real-time;
identifying in real-time the specific water-using appliance responsible for current
water usageby applying a trained appliance classification model to the received flow and
15 water level data;
predicting, using a multivariate time series forecasting model comprising a Long
Short-Term Memory (LSTM) neural network, a future water usage for each water-using
appliance based on historical usage patterns and ongoing tank water level changes; and
transmittingto a user device, the disaggregated appliance-level usage data and the
20 predicted future usage profile for enabling water conservation actions by a user.
8. The method as claimed in claim 7, wherein predicting future water consumption for each
water-using appliance comprises generating a time-stamped forecast of water usage for each
appliance for a predefined future time interval.
25
9. The method as claimed in claim 7, wherein disaggregated water usage for each appliance
is classified into activity categories comprising at least one of: filling, usage, and no activity,
and further wherein the machine learning model (112) generates confidence scores associated
with each classification.
30
10. The method as claimed in claim 7, wherein the machine learning model (112) is
periodically updated with newly labeled flow rate and water level data to continuously
improve the accuracy of appliance identification and usage prediction.
Dated this September 22nd , 2025
31
Arjun Karthik Bala (IN/PA 1021)
Agent for Applicant