Thermal power generating unit MPC prediction error optimization method based on stacked LSTM

Abstract

The invention discloses a thermal power generating unit MPC prediction error optimization method based on stacked LSTM. The method comprises the following steps: step 1, comprehensively collecting operation historical data of a thermal power generating unit under different working conditions; carrying out preprocessing operation on the collected original data; step 2, determining the number of layers of the stacked LSTM network and the number of LSTM neurons in each layer, establishing a basic network and initializing parameters; and step 3, training the stacked LSTM network, and performing error compensation on the prediction result of the standard MPC prediction model by using the trained stacked LSTM network. According to the method, accurate modeling and real-time compensation are carried out on MPC prediction errors by introducing a stacked LSTM architecture in deep learning, so that the control precision of the thermal power generating unit is remarkably improved, the adaptability and reliability of an MPC controller under complex working conditions are enhanced, the control effect of the thermal power generating unit is further improved, the operation cost is reduced, and stable operation of a power grid is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the field of thermal power generation control technology and relates to a method for optimizing the prediction accuracy of Model Predictive Control (MPC) using a deep learning method, and specifically relates to a method for optimizing the MPC prediction error of a thermal power unit based on stacked LSTM. Background Art

[0002] As core infrastructure for energy conversion, advances in thermal power generation control technology are of great strategic significance for improving energy efficiency, reducing operating costs, and ensuring stable grid operation. A thermal power unit is a typical large-scale, complex industrial process system, whose operation involves the coupling of multiple disciplines, including thermodynamics, fluid mechanics, and chemical reactions. Numerous physical and chemical variables in the system interact in complex ways, and these relationships change dynamically with changes in the unit's operating conditions.

[0003] Model Predictive Control (MPC), with its superior multivariable control performance, has become a mainstream approach for advanced control of thermal power plants. MPC uses an online rolling optimization process to determine the optimal control strategy based on predictions of future system behavior. However, due to the inherently strong nonlinear characteristics of thermal power plant systems and the dynamic variability of operating conditions, the mathematical models derived from traditional system identification methods often fail to accurately describe actual system behavior. This model mismatch problem severely limits the effective control of MPC controllers.

[0004] In recent years, deep learning technology has achieved breakthroughs in various fields. The LSTM architecture has been particularly outstanding in processing time series data. Its unique gating mechanism and sequence modeling capabilities provide an effective approach to solving time series prediction problems. In the field of thermal power plant control, LSTM can effectively capture long-term and short-term dependencies in time series data, and its hierarchical feature extraction capabilities help build more accurate prediction models. Furthermore, the residual connections and layer normalization design in LSTM provide strong support for the stability of deep network training, providing a new technical approach for optimizing prediction errors in MPC prediction for thermal power plants. Summary of the Invention

[0005] To address the problem of insufficient prediction accuracy in traditional MPC control of thermal power plants due to model mismatch, this paper provides a method for optimizing MPC prediction errors for thermal power plants based on stacked LSTMs. By incorporating the stacked LSTM architecture from deep learning, this method accurately models and compensates for MPC prediction errors in real time, significantly improving the control accuracy of thermal power plants and enhancing the adaptability and reliability of MPC controllers under complex operating conditions. This further enhances the control effectiveness of thermal power plants, reduces operating costs, and ensures stable power grid operation.

[0006] The purpose of the present invention is achieved through the following technical solutions: A stacked LSTM-based MPC prediction error optimization method for thermal power units includes the following steps: Step 1: Data processing: Step 1-1, data collection: Comprehensively collect historical operating data of thermal power units under different operating conditions; Step 1-2, data preprocessing: preprocessing the collected raw data; Steps 1-3, data division: divided into training set and validation set according to the time series relationship; Step 2: Network design: Determine the number of layers of the stacked LSTM network and the number of LSTM neurons in each layer, build the basic network and initialize the parameters; Step 3: Error compensation: Step 3-1, offline training: Use the training set data to train the stacked LSTM network; Step 3-2, online operation: Use the trained stacked LSTM network to compensate for the error of the prediction results of the standard MPC prediction model.

[0007] Compared with the prior art, the present invention has the following advantages: 1. Improved prediction accuracy: While maintaining the original MPC control logic, the stacked LSTM deep learning method is introduced to effectively capture the nonlinear characteristics and dynamic changes of the thermal power unit system, significantly improving the accuracy of the prediction model, thereby enhancing the performance of the MPC controller and better achieving optimized control of the thermal power unit.

[0008] 2. Adaptive Performance: The stacked LSTM model has sequence modeling capabilities and can adaptively learn prediction error patterns under different operating conditions. This means that when the operating characteristics of the thermal power unit change, the method can automatically adjust without human intervention, ensuring the accuracy and effectiveness of prediction error compensation, and enhancing the adaptability and reliability of MPC control during operation.

[0009] 3. Engineering practicality: This method optimizes the traditional MPC framework while retaining it, making full use of the hardware and software resources of the existing MPC control system without the need for large-scale transformation of the original system. DETAILED DESCRIPTION

[0010] The technical solution of the present invention is further described below, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0011] The present invention provides a method for optimizing the MPC prediction error of a thermal power unit based on a stacked LSTM, the method comprising the following steps: Step 1: Data processing: Step 1-1, Data Collection: Comprehensively collect the historical operating data of the thermal power unit under different operating conditions. The data sources include the unit's DCS, online monitoring system, and historical database. The data is required to cover all operating conditions from the minimum operating condition to the maximum operating condition that the unit can achieve. The collected data includes key control variables (including coal feed rate, air volume, and water feed rate) and controlled variables (including main steam pressure, water wall outlet temperature, and unit power).

[0012] Step 1-2, Data Preprocessing: Preprocess the collected raw data. The specific steps are as follows: Step 1-2-1: Clean the data to remove outliers, missing values, and noise. For missing values, interpolate the values ​​based on the values ​​of adjacent data points; for outliers, set a reasonable threshold to detect and correct them.

[0013] Step 1-2-2: Standardize the data.

[0014] Steps 1-3, Data Partitioning: Data is divided into training and validation sets based on time series relationships. The training set is used to train the stacked LSTM network to learn the prediction error patterns of the thermal power unit system; the validation set is used to evaluate the performance of the standard MPC prediction model, monitor its performance on new data, prevent overfitting, and ensure that the standard MPC prediction model has good generalization capabilities.

[0015] Step 2: Network design: Determine the number of layers in the stacked LSTM network and the number of LSTM neurons in each layer, build the basic network, and initialize the parameters. The specific steps are as follows: Step 2-1: Standard MPC prediction model construction: Step 2-1-1: Use traditional system identification methods to establish a basic state space model of the thermal power unit. This model is based on the historical operating data of the unit and is verified through parameter estimation and model verification steps to ensure that it can preliminarily describe the dynamic behavior of the thermal power unit. Step 2-1-2: Based on step 2-1-1, a standard MPC prediction model is constructed. This model takes the basic state space model as the core, combines the control objectives and constraints of the unit, and uses a rolling optimization algorithm to predict the future state of the thermal power unit, providing a benchmark for subsequent prediction error optimization.

[0016] Step 2-2, stacked LSTM network design: A stacked LSTM network structure was designed, consisting of multiple LSTM layers. Each LSTM layer contains a certain number of neurons. These neurons, through a unique gating mechanism (input gate, forget gate, and output gate), effectively control the flow of information, avoiding the vanishing gradient problem that traditional RNNs encounter when processing long sequences. This stacked structure enables the network to capture dependencies in time series data from different perspectives, while the feedforward neural network is responsible for extracting deep feature representations, further enhancing the model's expressive power and predictive accuracy.

[0017] Considering that the inertia time of a boiler is around 5 minutes and that of a steam turbine is around 10 seconds, the number of neurons in the first LSTM layer was set to 256 to accurately characterize the timing characteristics of the thermal power unit. The intermediate LSTM layers were stacked layer by layer, with the number of neurons decreasing in each layer (128 → 64 → 32 → 16). The final LSTM layer was fully connected. A regularization layer was added after each LSTM layer to prevent overfitting.

[0018] Step 3: Error compensation: Step 3-1, Offline training: Use the training set data to train the stacked LSTM network. The specific steps are as follows: Step 3-1-1: The error between the actual system response in the historical operating data and the MPC predicted value is used as the output of the training sample, and the corresponding input is the historical prediction error sequence and related operating condition parameters.

[0019] Step 3-1-2: Adjust the parameters of the stacked LSTM network through an optimization algorithm (such as Adam) so that the stacked LSTM network can accurately predict the prediction error of the MPC model.

[0020] Step 3-2, online operation: Use the trained stacked LSTM network to compensate for the error of the prediction results of the standard MPC prediction model. The specific steps are as follows: Step 3-2-1: Use the standard MPC forecasting model to perform basic forecasting and obtain preliminary future state forecast values.

[0021] Step 3-2-2: Input the current working condition data into the trained stacked LSTM network to obtain an estimate of the future prediction error.

[0022] Step 3-2-3: Superimpose the error estimate obtained in step 3-2-2 on the basic prediction result in step 3-2-1 to obtain a more accurate prediction result, thereby effectively compensating for the MPC prediction error and improving the control accuracy of the thermal power unit.

Claims

1. A stacked LSTM-based MPC prediction error optimization method for thermal power units, characterized by The method comprises the following steps: Step 1: Data processing: Step 1-1, data collection: Comprehensively collect historical operating data of thermal power units under different operating conditions; Step 1-2, data preprocessing: preprocessing the collected raw data; Steps 1-3, data division: divided into training set and validation set according to the time series relationship; Step 2: Network design: Determine the number of layers of the stacked LSTM network and the number of LSTM neurons in each layer, build the basic network and initialize the parameters; Step 3: Error compensation: Step 3-1, offline training: Use the training set data to train the stacked LSTM network; Step 3-2, online operation: Use the trained stacked LSTM network to compensate for the error of the prediction results of the standard MPC prediction model.

2. The MPC prediction error optimization method for thermal power units based on stacked LSTM according to claim 1 is characterized in that In step 1-1, the collected data include key control variables and controlled variables. The key control variables include coal feed rate, air volume, and water feed rate. The controlled variables include main steam pressure, water wall outlet temperature, and unit power.

3. The MPC prediction error optimization method for thermal power units based on stacked LSTM according to claim 1 is characterized in that The specific steps of steps 1-2 are as follows: Step 1-2-1: Perform data cleaning to remove outliers, missing values ​​and noise data; Step 1-2-2: Standardize the data.

4. The MPC prediction error optimization method for thermal power units based on stacked LSTM according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 2-1: Standard MPC prediction model construction: Step 2-1-1: Use traditional system identification methods to establish the basic state space model of the thermal power unit; Step 2-1-2: Based on step 2-1-1, construct a standard MPC prediction model; Step 2-2, stacked LSTM network design: A stacked LSTM network structure with multiple LSTM layers is designed. Each LSTM layer contains a certain number of neurons. The number of neurons in the first LSTM layer is determined according to the characteristics of the thermal power unit equipment. The intermediate LSTM layers are stacked layer by layer, with the number of neurons in each layer decreasing. The final LSTM layer is fully connected. A regularization layer is added after each LSTM layer to prevent overfitting.

5. The MPC prediction error optimization method for thermal power units based on stacked LSTM according to claim 1 is characterized in that The specific steps of step 3-1 are as follows: Step 3-1-1: The error between the actual system response in the historical operating data and the MPC predicted value is used as the output of the training sample, and the corresponding input is the historical prediction error sequence and related operating condition parameters; Step 3-1-2: Adjust the parameters of the stacked LSTM network through the optimization algorithm so that the stacked LSTM network can accurately predict the prediction error of the MPC model.

6. The method for optimizing the MPC prediction error of a thermal power plant based on stacked LSTM according to claim 1, characterized in that The specific steps of step 3-2 are as follows: Step 3-2-1: Use the standard MPC forecasting model to perform basic forecasting and obtain preliminary future state forecast values; Step 3-2-2: Input the current working condition data into the trained stacked LSTM network to obtain an estimate of the future prediction error; Step 3-2-3: Superimpose the error estimate obtained in step 3-2-2 on the basic prediction result in step 3-2-1 to obtain a more accurate prediction result, thereby effectively compensating for the MPC prediction error.