1. A method for enhancing energy forecasting in combined cycle power plants using a hybrid ConvLSTM and FC neural network model, wherein the method comprising: a data acquisition module that is configured to acquire real-time and/or historical operational and environmental parameter values from components of a plant, including gas turbines, heat recovery steam generators (HRSG), steam turbines, and auxiliary systems; a preprocessing module that is configured to preprocess the acquired multivariate time-series data to produce structured input tensors; and a predictive engine comprising a hybrid deep learning structure including at least one convolutional long short-term memory (ConvLSTM) layer and at least one fully connected (FC) neural network layer, wherein the at least one ConvLSTM layer is configured to extract spatiotemporal features from the structured input tensors, and the at least one FC neural network layer is configured to perform nonlinear regression to produce forecasted energy output values.

Description:ENHANCING ENERGY FORECASTING IN COMBINED CYCLE POWER PLANTS USING A HYBRID CONVLSTM AND FC NEURAL NETWORK MODEL

Technical Field
[0001] The embodiments herein generally relate to a method for enhancing energy forecasting in combined cycle power plants using a hybrid ConvLSTM and FC neural network model.
Description of the Related Art
[0002] The existing forecasting models that are used in CCPPs are based on statistical regression techniques and neural network models, which are not effective in handling nonlinear time dependencies, environmental variations, and load fluctuations. The primary object of the invention is to propose a highly accurate energy forecasting system using a combination of ConvLSTM and Fully Connected neural networks. Another object of the invention is to efficiently capture both spatial correlation and time dependency in the multi-variable operational parameters such as ambient temperature, exhaust vacuum, relative humidity, and atmospheric pressure.
[0003] A further object of the invention is to enhance the operational efficiency, reduce the energy consumption, minimize the forecasting error, and optimize the load scheduling techniques. An artificial intelligence-driven energy analytics and power system optimization, offering a computationally efficient real-time forecasting approach. In order to handle operational and environmental fluctuations, it addresses the use of a hybrid Convolutional Long Short-Term Memory (ConvLSTM) and Fully Connected (FC) neural network model for short-term and medium-term energy forecasting in Combined Cycle Power Plants (CCPPs).
SUMMARY
[0004] In view of the foregoing, an embodiment herein provides a method for enhancing energy forecasting in combined cycle power plants using a hybrid ConvLSTM and FC neural network model. In some embodiments, wherein a hybrid deep learning framework for precise energy output forecasting in Combined Cycle Power Plants that combines Fully Connected (FC) neural network layers with Convolutional Long Short-Term Memory (ConvLSTM) layers.
[0005] The ConvLSTM component efficiently models nonlinear relationships and dynamic environmental fluctuations by extracting spatial-temporal characteristics from sequential plant sensor data. For accurate regression-based energy prediction, the retrieved features are subsequently processed through fully linked layers. Data preprocessing, normalization, sliding window sequence generation, backpropagation model training, and performance assessment using RMSE and MAE metrics are all included in the system. Prediction accuracy, operational planning, and energy efficiency in CCPP operations are all much enhanced by the suggested hybrid architecture.
[0006] 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 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
[0007] The embodiments herein will be better understood from the following detailed description with reference to the drawings, in which:
[0008] FIG. 1 illustrate the method for enhancing energy forecasting in combined cycle power plants using a hybrid ConvLSTM and FC neural network model according to an embodiment herein.
DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS
[0009] 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 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 the scope of the embodiments herein.
[0010] FIG. 1 illustrate the method for enhancing energy forecasting in combined cycle power plants using a hybrid ConvLSTM and FC neural network model according to an embodiment herein. The preferred embodiments of the system and method for improving short-term and medium-term energy prediction of combined cycle power plants (CCPPs) through a combination of a Convolutional Long Short-Term Memory (ConvLSTM) neural network and a Fully Connected (FC) neural network. The preferred embodiments of the invention presented here offer a powerful, scalable, and accurate framework for modeling nonlinear relationships, as well as temporal and spatial relationships associated with combined cycle power generation systems.
[0011] In one particular embodiment, the system can be described as comprising a data acquisition module that can be configured for the collection of real-time and historical operational parameters from the combined cycle power plant. In the context of the combined cycle power plant, the gas turbines, the heat recovery steam generators, the steam turbines, the cooling systems, and the various other auxiliary systems can be included. At critical points, sensor arrays can be used for the continuous collection of various parameters, which can include the ambient temperature, the atmospheric pressure, the relative humidity, the exhaust vacuum, the fuel flow rate, the turbine inlet temperature, the compressor discharge pressure, the steam pressure, the steam temperature, the load demand, and the output. The data acquisition module can be designed for communication with the Supervisory Control and Data Acquisition systems, as well as the Distributed Control systems, via the Modbus, the OPC-UA, and the MQTT communication protocols.
[0012] In a preferred form, the raw data is subjected to a preprocessing step before it is fed into the model for training and prediction. The preprocessing block includes data cleaning, which removes noise, sensor anomalies, and handles missing data through interpolation or statistical imputation. Outlier detection is carried out through statistical, z-score, or isolation forest methods for robustness in the data. The data is then subjected to normalization or standardization through min-max scaling or z-score normalization for numerical stability during neural network training. Time-series data is segmented into segments, which represent a window of data over a specific time interval, e.g., 15 minutes, 30 minutes, or 1 hour, etc., as required for the process.
[0013] In the preferred embodiment, the core predictive engine employs a hybrid structure consisting of ConvLSTM layers and Fully Connected (FC) layers. The ConvLSTM module is designed to process spatiotemporal relationships within multivariate time series. Unlike the conventional LSTM network, the ConvLSTM network includes a convolutional structure within the gating mechanism to enable the extraction of local feature representations over time. Tensors are arranged as multidimensional arrays, with one dimension dedicated to temporal sequences and another dimension dedicated to feature channels corresponding to various plant parameters. The convolutional kernels enable the extraction of dynamic correlations between environmental factors and plant operational states.
[0014] In some embodiment, a variant of the proposed ConvLSTM structure is presented based on a stack of multiple ConvLSTM layers for hierarchical feature extraction. The input, forget, and output gates of each ConvLSTM layer are implemented through convolutional operations rather than matrix multiplication. The kernel sizes of the convolutional operations can range from 2×2 to 5×5 depending on the data resolution or granularity. The range of the number of filters of each layer can extend from 32 to 128 to achieve a trade-off between the complexity of the proposed structure and efficiency considerations. Dropout regularization and batch normalization are optionally included between the layers to prevent overfitting and ensure stable convergence of the proposed structure during training. The output of the proposed ConvLSTM structure is flattened into a feature vector of one dimension.
[0015] In the preferred embodiment, the feature vector flattened into a one-dimensional array is input into a Fully Connected (FC) Neural Network module. This FC module consists of one or more dense layers that are designed for nonlinear regression for the purpose of forecasting the power output. Rectified Linear Units (ReLU) are used for the activation functions for the hidden layers, while the output layer consists of a linear activation function for the purpose of predicting continuous output for the energy value.
[0016] In another embodiment, the hybrid model can be trained via the supervised learning paradigm. In this case, the labeled data set consists of historical operational data with the corresponding actual output energy value. The type of loss function that can be used for the purpose of training the model can be the Mean Squared Error (MSE) function, the Root Mean Squared Error (RMSE) function, the Mean Absolute Percentage Error (MAPE) function, and so on, depending upon the desired evaluation criteria. In this case, optimization algorithms such as Adam and RMSProp can be used for the purpose of optimizing the model via backpropagation through time (BPTT). In addition, the learning rate can be adjusted via adaptive learning rate optimization techniques.
[0017] In another embodiment, the system can be designed with a cross-validation feature that improves the capability for generalization. The data can be split into training, validation, and test sets, for example, with a split ratio of 70:15:15. K-fold cross-validation can be used for the robustness of the model with respect to data bias. Grid search can be used for the optimization of the model with respect to the kernel size, number of filters, number of neurons in the FC layer, learning rate, batch size, and dropout ratio.
[0018] In another embodiment, the system can be designed for both STF and DFA. In the case of STF, high-resolution real-time data can be used, while for DFA, aggregated data sequences for one hour or one day can be used. In a specific example, the deployment architecture may comprise an inference engine integrated with plant control systems. In this example, the trained hybrid ConvLSTM-FC model may be deployed in a server environment enabled with GPU acceleration or a cloud computing environment enabled to scale resources appropriately. Real-time data inputs may be fed to the model, and predicted power generation values may be obtained.
[0019] In a preferred embodiment, a feedback loop mechanism may be included in the system, wherein errors in the forecast are monitored in real time, and performance metrics may be computed in real time, including RMSE, MAPE, and R² score. In the event of errors in the predictions, beyond a certain threshold, the system may automatically perform a retraining of the model using the latest data available in the system, thereby enabling adaptive learning in the model to ensure reliable performance in the long term, considering the effects of seasons, equipment, and fuel quality.
[0020] In another embodiment, explainability techniques are provided by including the explainability features of the model by utilizing techniques such as SHAP (SHapley Additive explanations) or attention-weight visualization. This allows the operators to have trust in the model by understanding the important parameters affecting the output of the energy.
[0021] In another embodiment, cybersecurity is provided by including cybersecurity techniques for secure data transmission. This is done by including cybersecurity techniques such as encryption, authentication, and firewall mechanisms. This ensures the data is secure, and the model is not tampered with by unauthorized users. The model is designed in such a manner that it is extensible and can be connected with multiple combined cycle power plants. The model is designed in such a manner that it can handle multiple locations. The model can be integrated with the energy management systems and smart grid systems through the application programming interface.
[0022] In comparison to other statistical-based conventional methods like ARIMA or even the LSTM model alone, the proposed hybrid approach using the ConvLSTM-FC model provides better accuracy for the forecast. This is because the proposed model effectively handles the nonlinear spatiotemporal relationships between the environmental and operational variables. Therefore, the preferred embodiments of the present invention propose a technically advanced and commercially viable solution to improve the accuracy of energy forecast in combined cycle power plants. As discussed, the proposed solution using the ConvLSTM-FC neural network provides a better prediction stability, which results in the reduction of operational costs, fuel efficiency, and grid reliability. Although the preferred embodiments of the present invention propose a specific solution using the ConvLSTM-FC neural network, the depth of the network, activation functions, optimization functions, and infrastructure to deploy the proposed solution can be varied without changing the essence of the proposed solution.
, Claims:
I/We Claim:
1. A method for enhancing energy forecasting in combined cycle power plants using a hybrid ConvLSTM and FC neural network model, wherein the method comprising:
a data acquisition module that is configured to acquire real-time and/or historical operational and environmental parameter values from components of a plant, including gas turbines, heat recovery steam generators (HRSG), steam turbines, and auxiliary systems;
a preprocessing module that is configured to preprocess the acquired multivariate time-series data to produce structured input tensors; and
a predictive engine comprising a hybrid deep learning structure including at least one convolutional long short-term memory (ConvLSTM) layer and at least one fully connected (FC) neural network layer, wherein the at least one ConvLSTM layer is configured to extract spatiotemporal features from the structured input tensors, and the at least one FC neural network layer is configured to perform nonlinear regression to produce forecasted energy output values.