Method for predicting and optimizing load distribution of cogeneration unit
By combining the prediction and optimization framework of TCN-ECA-BiLSTM and TD3 reinforcement learning algorithm, the problems of low prediction accuracy and slow optimization response in load allocation of cogeneration units are solved, and efficient load allocation and low-carbon operation are achieved.
Patent Information
- Application Number
- CN202511886449.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-03-17
AI Technical Summary
Existing load allocation methods for combined heat and power (CHP) units suffer from low prediction accuracy, slow optimization response, and difficulty in adapting to dynamic operating conditions. Furthermore, prediction and optimization are disconnected and fail to form an integrated framework, resulting in insufficient optimization potential for both steam production coal consumption rate and power generation steam consumption rate.
By combining the TCN-ECA-BiLSTM prediction model with the TD3 reinforcement learning algorithm, and through an integrated framework of feature extraction, prediction and optimization, redundant features are jointly removed using Spearman's rank correlation coefficient and mutual information. The TCN-ECA-BiLSTM model is constructed for high-precision prediction, and dynamic optimization is performed using the TD3 reinforcement learning algorithm to achieve real-time adjustment of steam production coal consumption rate and power generation steam consumption rate.
It has achieved high-precision prediction and dynamic optimization of the load of cogeneration units, improved prediction accuracy and optimization response speed, reduced coal consumption and power generation costs, enhanced the system's adaptability, and met the low-carbon operation requirements of the "dual carbon" strategy.
Smart Images
Figure CN121682116A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of combined heat and power system scheduling, and particularly relates to a combined heat and power unit load distribution prediction and optimization method. BACKGROUND
[0002] Under the background of the "double carbon" strategy, combined heat and power (CHP) units, as core energy conversion equipment with both power supply and heat supply functions, have become a key support for building a low-carbon energy system due to their high primary energy utilization rate and low carbon emissions. However, the operation process of combined heat and power units has strong coupling and nonlinear characteristics: on the one hand, the units need to meet both power grid power generation demand and user heat supply demand, and the dynamic changes in medium-pressure extraction flow (for heat supply) and low-pressure exhaust flow (for power generation) will directly affect system energy efficiency; on the other hand, affected by factors such as load fluctuations, coal quality changes, and equipment aging, traditional load distribution methods based on static models (such as linear programming and particle swarm optimization) have low prediction accuracy, slow optimization response, and difficulty in adapting to dynamic operating conditions.
[0003] In the prior art, scholars have carried out a large amount of research in the field of CHP system prediction and optimization, but there are still obvious defects: in terms of prediction, although models such as CNN-LSTM and PSO-BP can capture some time sequence characteristics, they cannot effectively handle long-distance time sequence dependence and dynamic weight distribution of key features, resulting in large prediction errors for key parameters such as medium-pressure extraction and low-pressure exhaust; in terms of optimization, traditional optimization algorithms (such as genetic algorithms and ADMM algorithms) rely on manual setting of static parameters, lack predictive consideration of future loads, and are prone to local optimization in multi-unit coupling scenarios, resulting in insufficient optimization of steam coal consumption rate and power generation steam consumption rate; at the same time, existing research often separates prediction and optimization, and does not form a prediction-optimization integrated framework, making it difficult to meet the real-time scheduling needs of combined heat and power systems. SUMMARY
[0004] The present application aims to provide a combined heat and power unit load distribution prediction and optimization method. The present application combines the TCN-ECA-BiLSTM prediction model with the TD3 reinforcement learning algorithm model to form a prediction-optimization integrated framework. After feature extraction, the TCN-ECA-BiLSTM prediction model is used to strengthen the prediction input by taking advantage of its key feature reinforcement and bidirectional sequence modeling. The TD3 reinforcement learning algorithm model considers steam coal consumption rate and power generation steam consumption rate for distribution, achieving real-time and reliable dynamic optimization.
[0005] The technical solution of the present application is a combined heat and power unit load distribution prediction and optimization method, which is performed according to the following steps: Step S1: Obtain historical data of the cogeneration unit, group them in two working conditions of medium pressure extraction and low pressure exhaust, extract features through joint of Spearman rank correlation coefficient and mutual information, eliminate redundant features and low correlation features, and obtain a target data set for model training; Step S2: Construct a TCN-ECA-BiLSTM prediction model, which includes a time series convolution network layer, an efficient channel attention layer and a bidirectional long short-term memory network layer; divide the target data set into a training set, a validation set and a test set, input the TCN-ECA-BiLSTM prediction model for training until the prediction effect of the TCN-ECA-BiLSTM prediction model on the medium pressure extraction flow and the low pressure exhaust flow reaches a preset target; Step S3: Based on the historical data of the cogeneration unit, construct a boiler side steam generation coal consumption rate fitting function and a turbine side power generation steam consumption rate fitting function, set the state space, action space and reward function of the TD3 reinforcement learning algorithm model for each function to train; input the output of the TCN-ECA-BiLSTM prediction model as the state input of the TD3 reinforcement learning algorithm model, optimize the coal consumption distribution through the steam generation coal consumption rate fitting function, and then combine the power generation steam consumption rate fitting function to distribute the steam flow, to obtain the optimized steam generation coal consumption rate and power generation steam consumption rate; train the TD3 reinforcement learning algorithm model to convergence to obtain a TD3 reinforcement learning algorithm optimization model; Step S4: Input the real-time running data of the cogeneration unit into the TCN-ECA-BiLSTM prediction model to obtain the predicted values of the low pressure exhaust flow and the medium pressure extraction flow; input the predicted values into the TD3 reinforcement learning algorithm optimization model to output the boiler coal consumption distribution scheme and the turbine steam flow distribution scheme, to realize dynamic optimization distribution of the cogeneration unit load.
[0006] In the above-mentioned method for predicting and optimizing load distribution of a cogeneration unit, the sampling period of the historical data of the cogeneration unit in step S1 is 1-2 minutes, and the collection time span is at least 10 days.
[0007] In the above-mentioned method for predicting and optimizing load distribution of a cogeneration unit, the feature extraction in step S1 is to first eliminate redundant features with an absolute value of the Spearman rank correlation coefficient greater than a set value, then normalize the Spearman rank correlation coefficient and the mutual information coefficient of each feature and add them according to a set value, select a plurality of features with an added value in a set position as prediction features, and finally add historical time series features to obtain input features.
[0008] In the aforementioned method for predicting and optimizing load allocation in a combined heat and power (CHP) unit, the TCN-ECA-BiLSTM prediction model in step S2 includes a temporal convolutional network layer, an efficient channel attention layer, and a bidirectional long short-term memory (LSTM) network layer. The temporal convolutional network layer captures long-term temporal dependencies through causal convolution and dilated convolution. The efficient channel attention layer calculates channel weights through local cross-channel convolution to enhance the attention to key features. The bidirectional LSTM network layer learns the bidirectional temporal correlation of sequence data in parallel through forward and backward LSTM units.
[0009] In the aforementioned method for predicting and optimizing load distribution in a combined heat and power (CHP) unit, the boiler-side steam production coal consumption rate fitting function in step S3 is constructed based on the boiler efficiency calculated using the inverse balance method. The boiler efficiency equation is shown below: ; In the formula, η is the boiler efficiency, A, B, and C are polynomial coefficients, and D... b Steam production of the boiler; The formula for the coal consumption rate for steam production is shown below: ; In the formula, b is the coal consumption rate for steam production, and n is the number of boilers. Let i be the steam output of the i-th boiler. This refers to the enthalpy value of the boiler's main steam. Let be the enthalpy of the boiler feedwater, h be the boiler efficiency function, and Q be the lower heating value of the coal.
[0010] In the aforementioned method for predicting and optimizing load allocation of a combined heat and power (CHP) unit, the turbine-side power generation steam consumption rate fitting function in step S3 is constructed based on the power generation after extraction and exhaust allocation using the random forest algorithm. The turbine power generation equation is shown below: ; In the formula, N is the power generation of the steam turbine, Gp is the medium-pressure extraction flow rate, Gg is the low-pressure exhaust flow rate, and g is the fitting function; The formula for the power generation steam consumption rate is: ; In the formula, d is the power generation steam consumption rate, m is the number of steam turbines, G is the total steam turbine intake, and P is the total steam turbine intake. i Let be the power output of the i-th steam turbine.
[0011] In the aforementioned method for predicting and optimizing load allocation of a combined heat and power (CHP) unit, when the optimization objective in step S3 is the steam production coal consumption rate, the state space of the TD3 reinforcement learning algorithm model includes the current coal consumption and efficiency of each boiler, the action space is the coal consumption allocation of each boiler, and the constraints are: ; The reward function is: ; In the formula, This represents the minimum coal feed rate for the i-th boiler. Let b be the maximum coal feed rate of the i-th boiler, and b be the optimized coal consumption rate for steam production.
[0012] In the aforementioned method for predicting and optimizing load allocation of a combined heat and power (CHP) unit, when the optimization objective in step S3 is the power generation steam consumption rate, the state space of the TD3 reinforcement learning algorithm model includes the intermediate-pressure extraction flow rate and the low-pressure exhaust flow rate of each turbine, and the action space includes the adjustment amount of the intermediate-pressure extraction flow rate and the low-pressure exhaust flow rate of each turbine. The constraints are as follows: ; ; ; In the formula, Let be the minimum generating capacity of the i-th steam turbine. Let be the maximum generating capacity of the i-th steam turbine; The maximum medium-pressure extraction flow rate of the i-th steam turbine; Let be the minimum low-pressure exhaust flow rate of the i-th steam turbine; The reward function is: ; In the formula, d represents the optimized power generation steam consumption rate.
[0013] In the aforementioned method for predicting and optimizing load allocation for combined heat and power (CHP) units, the optimized allocation in step S4 satisfies the following constraints: ; ; ; In the formula, To optimize the coal consumption of the i-th boiler, To optimize the coal consumption of the i-th boiler, To optimize the low-pressure exhaust flow of the i-th turbine, To optimize the low-pressure exhaust flow of the i-th steam turbine, To optimize the intermediate-pressure extraction flow rate of the i-th steam turbine, To optimize the intermediate-pressure extraction flow rate of the i-th steam turbine.
[0014] Compared with existing technologies, this invention uses a feature extraction method combining Spearman's rank correlation coefficient and mutual information to perform redundancy removal and key feature extraction on high-dimensional operating data of cogeneration units. This solves the problem of model overfitting caused by feature redundancy in traditional data preprocessing and avoids interference from low-correlation features on prediction accuracy. By constructing a TCN-ECA-BiLSTM prediction model, it can accurately predict the medium-pressure extraction flow and low-pressure exhaust flow of cogeneration units. The TCN-ECA-BiLSTM prediction model integrates the ability to capture long-term time-series dependencies and the key feature enhancement effect of efficient channel attention. Combined with the advantages of bidirectional sequence modeling, it accurately mines the complex nonlinear relationship between steam flow and unit operating parameters. Using the output of the TCN-ECA-BiLSTM prediction model as input, and combining it with the TD3 reinforcement learning algorithm model to construct an optimization framework, under the premise of constraining the total coal consumption of boilers and the total demand flow of medium and low-pressure steam, it adjusts the coal consumption allocation of each boiler and the steam flow allocation of each turbine to achieve real-time and reliable dynamic optimization. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a low-pressure exhaust flow characteristic extraction diagram of the present invention; Figure 3 This is a feature extraction diagram of the medium-pressure air extraction flow rate of the present invention; Figure 4 This is a fitting graph of boiler steam production and boiler efficiency according to the present invention; Figure 5 This is a comparison chart of the predicted and actual values of the low-pressure air extraction flow rate in the verification set of this invention. Figure 6 This is a comparison chart of the predicted and actual values of the centralized pressure exhaust flow rate for verification purposes in this invention. Figure 7 This is a comparison chart showing the coal consumption rate of the steam generation unit before and after optimization, serving as a verification of the present invention. Figure 8 This is a comparison chart showing the power generation steam consumption rate before and after optimization, serving as a verification of the present invention. Detailed Implementation
[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.
[0017] Example: A method for predicting and optimizing load allocation in combined heat and power (CHP) units, as shown in the attached document. Figure 1 As shown, proceed with the following steps: Step S1: Collect historical data of the cogeneration unit through the unit's SCADA system. The data collection object is the high-temperature ultra-high pressure circulating fluidized bed boiler with rated evaporation capacity. The sampling period is 1 minute, and the collection time span is no less than 10 days. The data is grouped into two operating conditions: medium-pressure extraction and low-pressure exhaust. Feature extraction is performed by combining Spearman's rank correlation coefficient and mutual information. Redundant features and low-correlation features are removed to obtain the target dataset for model training. Feature extraction involves first removing redundant features with Spearman rank correlation coefficients greater than 0.9, then normalizing the Spearman rank correlation coefficients and mutual information coefficients of each feature, and summing them up according to their respective 50% weightings, as shown in the appendix. Figure 2 and attached Figure 3 As shown, the features with the top 10 summed scores are used as prediction features. Finally, historical time series features (for example, if the prediction is medium-pressure pumping, the historical time series features formed by the medium-pressure pumping in the previous 15 time steps are also used as input features) are added to obtain 11-dimensional input features.
[0018] Step S2: Construct the TCN-ECA-BiLSTM prediction model. Divide the target dataset into training, validation and test sets in an 8:1:1 ratio. Input the dataset into the TCN-ECA-BiLSTM prediction model for training. Adjust the hyperparameters, including sliding window, convolution kernel size and dropout ratio, until the TCN-ECA-BiLSTM prediction model achieves the preset target in predicting medium-pressure pumping flow and low-pressure exhaust flow. The TCN-ECA-BiLSTM prediction model comprises a Temporal Convolutional Network (TCN), an Efficient Channel Attention (ECA) layer, and a Bidirectional Long Short-Term Memory (BiLSTM) network layer. The TCN layer captures long-term temporal dependencies through causal convolution and dilated convolution. Its convolutional channel counts are 32, 128, and 256, respectively, with a kernel size of 7. The Efficient Channel Attention layer calculates channel weights through local cross-channel convolutions, enhancing the attention to key features. The Bidirectional Long Short-Term Memory (BiLSTM) network layer models bidirectional temporal associations by processing sequential data in parallel using forward and backward Long Short-Term Memory units. The TCN-ECA-BiLSTM prediction model is trained using the Adam optimizer with a learning rate of 1×10⁻⁶. −5 The batch size is 64, the number of training epochs is 100, the Dropout ratio is 0.3, and the sliding window length is 15, as shown in Table 1 below: Parameter Name Value Explanation Sliding Window 15 Historical input step size of time series Batch size 64 Training batch size Input dimension 11 Number of input features Output dimension 1 Dimension of predicted variable Number of channels [32,128,256] Number of convolutional layers in each TCN layer Convolution kernel size 7 Convolution kernel size Dropout ratio 0.3 Regularization ratio Number of training rounds 100 Total number of training iterations Learning rate 1*10-5 Optimizer step size Table 1 Step S3: Based on historical data from the cogeneration unit, construct fitting functions for boiler-side steam production coal consumption rate and turbine-side power generation steam consumption rate. For each function, set the corresponding state space, action space, and reward function for the TD3 reinforcement learning algorithm model and train it. The TD3 reinforcement learning algorithm model is based on the Actor-Critic framework, a deep reinforcement learning algorithm including an actor network (main network), a target actor network, two critic networks (main network), and two target critic networks. Through the dual critic and minimum value mechanism, it effectively suppresses the inflation of Q-value estimation, making the agent's judgment of action value more realistic and reliable, and delaying updates. The combined mechanism with the target network ensures a smooth training process. Due to more accurate value estimation and more stable training, the TD3 reinforcement learning algorithm model can more effectively utilize collected empirical data (stored in the experience replay cache) for learning, learning better policies with less interaction data. Using the output of the TCN-ECA-BiLSTM prediction model as the state input of the TD3 reinforcement learning algorithm model, coal consumption allocation is optimized through the steam production coal consumption rate fitting function, and steam flow is allocated by combining it with the power generation steam consumption rate fitting function, resulting in optimized steam production coal consumption rate and power generation steam consumption rate. The TD3 reinforcement learning algorithm model is trained until convergence, yielding the optimized TD3 reinforcement learning algorithm model. The boiler-side steam production coal consumption rate fitting function is constructed based on the boiler efficiency calculated using the inverse balance method, as shown in the attached figure. Figure 4 As shown, the boiler efficiency equation is as follows: ; In the formula, η is the boiler efficiency, A, B, and C are polynomial coefficients, and D... b Steam production of the boiler; The formula for the coal consumption rate for steam production is shown below: ; In the formula, b is the coal consumption rate for steam production, and n is the number of boilers. Let i be the steam output of the i-th boiler. This refers to the enthalpy value of the boiler's main steam. h(G) represents the enthalpy of boiler feedwater. bi ) is the boiler efficiency function, and Q is the lower heating value of coal (taken as 29307.6 kJ / kg).
[0019] The turbine-side power generation steam consumption rate fitting function is constructed based on the power generation power after extraction and exhaust allocation using the random forest algorithm. The turbine power generation power equation is shown below: ; In the formula, N is the power generation of the steam turbine (MW), Gp is the medium-pressure extraction flow rate (t / h), Gg is the low-pressure exhaust flow rate (t / h), and g is the fitting function; The formula for the power generation steam consumption rate is: ; In the formula, d is the power generation steam consumption rate, m is the number of steam turbines, G is the total steam turbine intake (t / h), and P is the total steam turbine intake. i Let be the power generation capacity (MW) of the i-th steam turbine.
[0020] When the optimization objective is the coal consumption rate for steam production, the state space of the TD3 reinforcement learning algorithm model includes the current coal consumption and efficiency of each boiler, the action space is the coal consumption allocation for each boiler, and the constraints are: ; The reward function is: ; In the formula, This represents the minimum coal feed rate for the i-th boiler. Let b be the maximum coal feed rate of the i-th boiler, and b be the optimized coal consumption rate for steam production.
[0021] When the optimization objective is the power generation steam consumption rate, the state space of the TD3 reinforcement learning algorithm model includes the intermediate-pressure extraction flow rate and the low-pressure exhaust flow rate of each turbine, and the action space includes the adjustment amount of the intermediate-pressure extraction flow rate and the low-pressure exhaust flow rate of each turbine. The constraints are as follows: ; ; ; In the formula, Let be the minimum generating capacity of the i-th steam turbine. Let be the maximum generating capacity of the i-th steam turbine; The maximum medium-pressure extraction flow rate of the i-th steam turbine; Let be the minimum low-pressure exhaust flow rate of the i-th steam turbine; The reward function is: ; In the formula, d represents the optimized power generation steam consumption rate.
[0022] Step S4: Input the real-time operating data of the cogeneration unit into the TCN-ECA-BiLSTM prediction model to obtain the predicted values of low-pressure exhaust flow and medium-pressure extraction flow. Validate the predicted values against the actual values for the first 200 minutes. Figure 5 and Figure 6 As shown; the predicted values are input into the TD3 reinforcement learning algorithm optimization model, which outputs boiler coal consumption allocation scheme and turbine steam flow allocation scheme, realizing dynamic optimization allocation of load for cogeneration units. The comparison results are shown in the appendix. Figure 7 and attached Figure 8 As shown; Optimal allocation satisfies the following constraints: ; ; ; In the formula, To optimize the coal consumption of the i-th boiler, To optimize the coal consumption of the i-th boiler, To optimize the low-pressure exhaust flow of the i-th turbine, To optimize the low-pressure exhaust flow of the i-th steam turbine, To optimize the intermediate-pressure extraction flow rate of the i-th steam turbine, To optimize the intermediate-pressure extraction flow rate of the i-th steam turbine.
[0023] The TCN-ECA-BiLSTM prediction model's performance on the test set for predicting low-pressure exhaust flow rate is: R²=0.998, MAE=2.43t / h, RMSE=3.66t / h, MAPE=0.84%. The same model's performance is: R²=0.979, MAE=0.8t / h, RMSE=1.2t / h, MAPE=1.37%. See Table 2 below.
[0024] Table 2 The average optimization side represents the average of all key parameters before and after optimization across all points in the validation set. The average total coal consumption of the three boilers remains unchanged, while the average total steam flow increases by 2.55 t / h, and the average steam production coal consumption rate decreases by 0.45 kg coal / t steam. The two steam turbines increase their average hourly capacity by 5.17 MW, and the average power generation steam consumption rate decreases by 0.457 t / MWh. Based on an annual operating time of 6000 h and an electricity price of approximately RMB 0.538 / kWh in the first half of 2025, this would reduce carbon dioxide emissions by 3241.19 tons annually if the same amount of steam is produced. With the total medium-pressure steam flow and low-pressure steam demand remaining constant, the power generation revenue increases by RMB 16,688,760, while the total coal consumption remains unchanged; therefore, the power plant's annual net profit is RMB 16,688,760.
[0025] In summary, the load allocation prediction and optimization method for cogeneration units proposed in this invention is based on the construction of an integrated technical framework of feature extraction, accurate prediction, and dynamic optimization. This effectively solves the pain points of traditional methods, such as low prediction accuracy, slow optimization response, and susceptibility to local optima, demonstrating significant technological innovation and practical value. In the feature processing stage, a joint extraction strategy of Spearman's rank correlation coefficient and mutual information is adopted. Highly redundant features are first removed, then highly correlated key features are extracted and supplemented with core historical time-series features. This solves the problem of model overfitting caused by redundant information in high-dimensional operating data and avoids the interference of low-correlation features on prediction accuracy. It provides a high-quality, low-dimensional target dataset for subsequent model training, balancing data effectiveness and computational efficiency. In the prediction module, the TCN-ECA-BiLSTM prediction model achieves complementary advantages of multiple networks. The temporal convolutional network layer accurately captures long-term temporal dependencies through causal convolution and dilated convolution. The efficient channel attention layer dynamically allocates channel weights and strengthens the attention to key features through local cross-channel convolution. The bidirectional long short-term memory network layer models bidirectional temporal correlations by processing sequential data in bidirectional parallel. After hyperparameter optimization, the model achieves an R² of 0.998 and a MAPE of only 0.84% for low-pressure exhaust flow and an R² of 0.979 and a MAPE of 1.37% for medium-pressure pumping flow. The prediction accuracy is better than traditional CNN-LSTM, PSO-BP and other models, providing reliable data support for subsequent optimization. In the optimization phase, the method constructs fitting functions for boiler-side steam production coal consumption rate and turbine-side power generation steam consumption rate based on historical data. The boiler-side function is constructed based on boiler efficiency calculated using the inverse balance method, while the turbine-side function is based on power generation modeling using the random forest algorithm, ensuring the accurate quantification of the optimization objectives. Simultaneously, the TD3 reinforcement learning algorithm model is employed, designing reasonable state spaces, action spaces, constraints, and reward functions for the two core objectives of steam production coal consumption rate and power generation steam consumption rate. Using the output of the prediction model as the state input, refined training of coal consumption allocation and steam flow allocation is achieved. Ultimately, under the constraint of keeping the total boiler coal consumption and the total demand flow of medium and low-pressure steam constant, dynamic optimization allocation is achieved through a dual-commentator network and a delayed policy update mechanism. Furthermore, the integrated framework achieves deep fusion of prediction and optimization, with fast response speed and strong adaptability, effectively coping with dynamic operating conditions such as load fluctuations, coal quality changes, and equipment aging. This changes the current situation where traditional static optimization methods are difficult to adapt to real-time scheduling requirements, providing reliable technical support for the low-carbon and efficient operation of cogeneration systems under the "dual-carbon" strategy.
Claims
1. A method for cogeneration unit load dispatch forecasting and optimization, characterized in that: The following steps are taken: Step S1: Obtain historical data of the combined heat and power unit, group it under two working conditions of medium pressure extraction and low pressure exhaust, and extract features through the joint of Spearman rank correlation coefficient and mutual information, eliminate redundant features and low correlation features, and obtain a target data set for model training; Step S2: Construct a TCN-ECA-BiLSTM prediction model, which includes a time series convolutional network layer, an efficient channel attention layer, and a bidirectional long short-term memory network layer; divide the target data set into a training set, a validation set, and a test set, and input it into the TCN-ECA-BiLSTM prediction model for training until the prediction effect of the TCN-ECA-BiLSTM prediction model on the medium pressure extraction flow and the low pressure exhaust flow reaches a preset target; Step S3: Based on the historical data of the combined heat and power unit, construct a boiler side steam generation coal consumption rate fitting function and a steam turbine side power generation steam consumption rate fitting function, and set the state space, action space and reward function of the TD3 reinforcement learning algorithm model for each function to train; input the output of the TCN-ECA-BiLSTM prediction model as the state input of the TD3 reinforcement learning algorithm model, optimize the coal consumption distribution through the steam generation coal consumption rate fitting function, and then distribute the steam flow combined with the power generation steam consumption rate fitting function to obtain the optimized steam generation coal consumption rate and power generation steam consumption rate; train the TD3 reinforcement learning algorithm model to convergence to obtain a TD3 reinforcement learning algorithm optimization model; Step S4: Input the real-time running data of the combined heat and power unit into the TCN-ECA-BiLSTM prediction model to obtain the predicted values of the low pressure exhaust flow and the medium pressure extraction flow; input the predicted values into the TD3 reinforcement learning algorithm optimization model to output the boiler coal consumption distribution scheme and the steam turbine steam flow distribution scheme, and realize dynamic optimization distribution of the combined heat and power unit load.
2. The method of claim 1, wherein: The sampling period of the combined heat and power unit historical data in step S1 is 1-2 minutes, and the collection time span is at least 10 days.
3. The method of claim 1, wherein: The feature extraction in step S1 is to first eliminate redundant features with an absolute value of the Spearman rank correlation coefficient greater than a set value, then normalize the Spearman rank correlation coefficient and the mutual information coefficient of each feature and add them according to a set score, select multiple features with an added score in a set position as prediction features, and finally add historical time series features to obtain input features.
4. The method of claim 1, wherein: The TCN-ECA-BiLSTM prediction model of step S2 includes a time series convolutional network layer, an efficient channel attention layer, and a bidirectional long short-term memory network layer; the time series convolutional network layer captures long time series dependencies through causal convolution and dilated convolution; The efficient channel attention layer calculates channel weights through local cross-channel convolution to enhance the attention of key features; the bidirectional long short-term memory network layer learns the bidirectional time series association of sequence data through forward and reverse LSTM units in parallel.
5. The method of claim 1, wherein: The boiler side steam generation coal consumption rate fitting function in step S3 is constructed based on the boiler efficiency calculated by the counterbalance method, and the boiler efficiency equation is as follows: ； wherein η is the boiler efficiency, A, B, C are polynomial coefficients, D b is the steam production of the boiler; The steam generation coal consumption rate formula is as follows: ； where b is the steam generation coal consumption rate, n is the number of boilers, is the steam generation of the i-th boiler, is the enthalpy of the main steam of the boiler, is the enthalpy of the boiler feed water, h is the boiler efficiency function, and Q is the low heat value of the coal.
6. The combined heat and power plant load dispatch prediction and optimization method of claim 1, wherein: The steam turbine side power generation steam consumption rate fitting function in the step S3 is constructed based on the steam extraction and exhaust distribution of the random forest algorithm, and the steam turbine power generation power equation is as follows: ； In the formula, N is the steam turbine power generation power, Gp is the medium-pressure steam extraction flow, Gg is the low-pressure exhaust flow, and g is the fitting function. The power generation steam consumption rate formula is: ； In the formula, d is the power generation steam consumption rate, m is the number of steam turbines, G is the total steam turbine intake, P i is the power generated by the i-th steam turbine.
7. The combined heat and power plant load dispatch prediction and optimization method of claim 1, wherein: When the optimization target in the step S3 is the steam generation coal consumption rate, the state space of the TD3 reinforcement learning algorithm model includes the current coal consumption of each boiler and the efficiency of each boiler, the action space is the coal consumption distribution of each boiler, and the constraint condition is: ； The reward function is: ； In the formula, is the minimum coal supply amount of the i-th boiler, is the maximum coal supply amount of the i-th boiler, and b is the optimized steam generation coal consumption rate.
8. The combined heat and power plant load dispatch prediction and optimization method of claim 1, wherein: When the optimization target in the step S3 is the power generation steam consumption rate, the state space of the TD3 reinforcement learning algorithm model includes the medium-pressure steam extraction flow of each steam turbine and the low-pressure exhaust flow of each steam turbine, the action space is the medium-pressure steam extraction flow and low-pressure exhaust flow adjustment amount of each steam turbine, and the constraint condition is: ； ； ； wherein Pmin,i is the minimum power generated by the i-th steam turbine, Pmax,i is the maximum power generated by the i-th steam turbine; Qmax,i is the maximum medium-pressure extraction flow rate of the i-th steam turbine; Qmin,i is the minimum low-pressure exhaust flow rate of the i-th steam turbine. The reward function is: ； In the formula, d is the optimized power generation steam consumption rate.
9. The combined heat and power plant load dispatch prediction and optimization method of claim 1, wherein: The optimization distribution in the step S4 satisfies the following constraint conditions: ； ； ； wherein the coal consumption of the i-th boiler before optimization, the coal consumption of the i-th boiler after optimization, the low-pressure exhaust flow of the i-th steam turbine before optimization, the low-pressure exhaust flow of the i-th steam turbine after optimization, the medium-pressure extraction flow of the i-th steam turbine before optimization, the medium-pressure extraction flow of the i-th steam turbine after optimization.