Thermal power generating unit key parameter advanced prediction method based on physical information Transform

Abstract

The invention relates to the field of advanced prediction of key parameters of a coal-fired thermal power generating unit, in particular to an advanced prediction method for key parameters of a thermal power generating unit based on physical information Transform. Comprising the following steps: S1, acquiring historical operation data of a unit, and dividing the historical operation data of the unit into time sequence samples with labels; s2, encoding the time sequence sample into a feature containing a long-term historical state and input information of the unit for representation; s3, carrying out advanced prediction on key parameters of the unit; s4, on the basis of the thermal power generating unit mechanism model, constructing a physical information regularization item, and obtaining a loss function with physical information regularization; s5, training a thermal power generating unit key parameter advanced prediction Transform model through a mean square error loss function with a physical information regularization item; and S6, obtaining online real-time data of the unit, inputting the trained thermal power generating unit key parameter advanced prediction Transform model, and obtaining an advanced prediction result of the unit key parameters.

Description

Technical Field

[0001] The invention relates to the field of advance prediction of key parameters of coal-fired thermal power units, and in particular to an advance prediction method of key parameters of thermal power units based on physical information Transformer. Background Art

[0002] In the energy revolution of large-scale access of new energy power to the grid, coal-fired thermal power units have gradually transformed into regulating and supporting power sources. In order to give full play to the basic guarantee and system regulation functions, meet the requirements of stable operation of the grid and consumption of new energy power, coal-fired thermal power units generally need to participate in deep peak regulation, showing the operating characteristics of large-scale rapid load changes.

[0003] In order to achieve rapid load changes to smooth out fluctuations in grid demand and ensure economical and low-carbon operation of coal-fired power units, the flexibility and accuracy of unit control are crucial. Boilers are complex systems with large inertia and large lags. To achieve accurate and flexible control of them, it is necessary to achieve multi-step advance prediction of key parameters of thermal power units, adjust control instructions in advance, and ensure flexible, stable and efficient operation of thermal power units.

[0004] Under the deep peak load regulation and large-scale load changes, the unit operation presents different modal characteristics. How to achieve multi-step advance prediction of key parameters under all operating conditions is a difficult problem. Although the traditional mechanism modeling method of thermal power units is based on physical laws, it better reflects the dynamic characteristics of the unit and is generalizable for different operating conditions, but there are the following problems: 1) The mechanism of thermal power units is complex, and mechanism modeling is usually based on certain assumptions and simplifications, which cannot guarantee the modeling accuracy under all operating conditions; 2) The mechanism model of thermal power units has a large computational burden in the solution process and low multi-step prediction efficiency; 3) The mechanism modeling does not fully consider the impact of the long-term historical input of the system on the future system parameters, and there are unmodeled dynamic characteristics.

[0005] The data-driven approach directly learns the rules for predicting key parameters from the unit operation data, which is time-consuming in the offline model training phase in exchange for high efficiency in the online prediction phase. However, the prediction effect of the data-driven approach is highly dependent on the distribution characteristics of the training data, and accuracy can only be guaranteed when the modeling data and online real-time data meet the same probability distribution. Faced with a thermal power unit system with frequent data distribution changes under deep peak load regulation and large-scale load changes, the data-driven model has the risk of mismatch, which affects the feasibility of actual deployment.

[0006] In order to solve the problem of multi-step advance prediction of key parameters of thermal power units under deep peak regulation, avoid model mismatch under frequent changes in data distribution, reduce the high dependence of data-driven models on the quantity and quality of modeling data, and avoid the computational burden and accuracy loss of mechanism modeling and simulation, it is necessary to combine mechanism modeling and data-driven modeling, utilize physical information machine learning and the long-distance dependency capture ability of Transformer networks, train neural networks to learn the dynamic characteristics of units and the physical laws in mechanism models, and achieve high data efficiency and high generalization of advance prediction of key parameters of thermal power units. Summary of the invention

[0007] In order to solve the above problems, the present invention provides a method for advanced prediction of key parameters of thermal power units based on physical information Transformer.

[0008] The present invention adopts the following technical solution: a method for advanced prediction of key parameters of thermal power units based on physical information Transformer, comprising: S1: Obtain the historical operation data of the unit and divide the historical operation data of the unit into labeled time series samples; S2: Encode the time series samples into features containing the long-term historical state and input information of the unit for representation; S3: Advance prediction of key parameters of the unit; S4: Based on the mechanism model of thermal power units, a physical information regularization term is constructed to obtain a loss function with physical information regularization; S5: Train the Transformer model for advanced prediction of key parameters of thermal power units through the mean square error loss function with physical information regularization term; S6: Obtain the online real-time data of the unit, input the trained Transformer model for advanced prediction of key parameters of the thermal power unit, and obtain the advanced prediction results of the key parameters of the unit.

[0009] In some embodiments, in step S1, for the historical data of the unit with a time length of T and a sampling period of Δt, each time step A training sample can be extracted, where p is the length of historical data in each sample and f is the number of forward prediction steps; by t -p to t The unit load, drum pressure, main steam pressure, coal supply, and valve opening of the step are taken as a time series sample X t =[x t-p ,x t-p+1 ,…,x t ]; t +1 to t+f step unit load, drum pressure, main steam pressure as the corresponding label Y t , t arrive t +f-1 step unit load, drum pressure, main steam pressure, coal supply, and regulating valve opening are used as auxiliary variables A for calculating physical information constraints. t , build a labeled training dataset .

[0010] In some embodiments, step S2 includes: There will be labeled training samples X t Input Transformer encoder, extract long-distance dependencies to form feature representation z containing the long-term historical state of the unit and input information N .

[0011] In some embodiments, step S3 includes: The feature representation of long-term historical information is z N The input Transformer decoder makes advance predictions on the key parameters of the unit for the next f steps.

[0012] In some embodiments, in step S4, the mechanism model of the thermal power unit is: Among them, k2 and k3 are normalization coefficients, t e It is the time constant of the turbine valve adjustment process. T is the total volume of the working fluid in each stage of superheater, ρ T is the steam density in the superheater, C0 is the damping coefficient in the superheater, ρ D is the density of saturated steam accumulated in the drum, V s is the volume of the gas phase in the drum, Q net,ar The low calorific value of the coal entering the furnace. h s is the specific enthalpy of saturated steam, h fw is the specific enthalpy of feed water at the evaporation system inlet, and q4 is the heat loss on the boiler side.

[0013] In some embodiments, in step S4, the loss function is: in, To constrain the prediction sequence, Y t is the true value, S is the total number of training samples; , , yes The advanced prediction values ​​of medium load, drum pressure and main steam pressure, and Yes A t Medium coal feeding amount and regulating valve opening value, , and Yes t The true values ​​of medium load, drum pressure and main steam pressure, k2 and k3 are normalization coefficients, t e It is the time constant of the turbine valve adjustment process. T is the total volume of the working fluid in each stage of superheater, ρ T is the steam density in the superheater, C0 is the damping coefficient in the superheater, ρ D is the density of saturated steam accumulated in the drum, V s is the volume of the gas phase in the drum, Q net,ar The low calorific value of the coal entering the furnace. h s is the specific enthalpy of saturated steam, h fw is the specific enthalpy of feed water at the evaporation system inlet, and q4 is the heat loss on the boiler side.

[0014] In some embodiments, the Transformer model for advanced prediction of key parameters of thermal power units includes: An encoder, the encoder comprising an input embedding layer and N stacked self-attention layers, the encoder takes historical time series data of operating parameters such as unit load, drum pressure, main steam pressure, coal feed rate, and throttle opening as input, extracts long-distance dependencies through the embedding layer and the self-attention layer, and forms a feature representation containing the long-term historical state of the unit and input information; The decoder includes an input embedding layer, M stacked mutual attention layers and an output linear transformation layer. The decoder takes the historical time series of the key parameters to be predicted and the future time series initialized to 0 as input, combines the feature representation output by the encoder, and predicts the future time series through the embedding layer, the mutual attention layer and the output layer.

[0015] In some embodiments, step S5 includes: S51: Set the training set D={X t ,Y t ,A t} training sample Xt Input the Transformer advance prediction model to obtain the advance prediction results of the key parameters of the unit; S52: Calculate the advance prediction result and label Y t The mean square error loss between the forward prediction results and the auxiliary variable A t and system parameters are used to calculate the value of the physical information regularization term; Multiply the physical information regularization term by the weight λ and add it to the mean square error true value loss to get the total loss function value; S53: Input the total loss function value into the Adam optimizer to update the parameters of the Transformer model; S54: Repeat steps S51-S53 until the total loss function value basically stops decreasing, and finally obtain the Transformer advance prediction model with the best prediction effect.

[0016] In some embodiments, step S6 includes: S61: Deploy the Transformer model for predicting the key parameters of the thermal power unit trained in the offline stage to the computer on the unit site, and obtain the online real-time data of the unit through the DCS system, including the unit load, drum pressure, main steam pressure, coal supply, and regulating valve opening in the past p steps; S62: The unit load, drum pressure, main steam pressure, coal feed rate, and regulating valve opening of the past p steps are input into the Transformer model for advanced prediction of key parameters of the thermal power unit to obtain the advanced prediction results of the key parameters of the unit.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a method for advanced prediction of key parameters of coal-fired thermal power units based on physical information machine learning and Transformer neural network. The Transformer encoder based on the self-attention mechanism extracts the long-term dependency of key parameters such as coal-fired unit load, drum pressure, main steam pressure, and historical time series of control inputs such as coal feed rate and throttle opening to form a characteristic representation of the unit state; based on the long-term characteristic representation of the unit state, the Transformer decoder based on the mutual attention mechanism is used to predict key parameters such as unit load, drum pressure, main steam pressure, etc. in the future multiple steps; the mechanism model of the unit is discretized by the Euler method to construct a physical information regularization term, and the training of the Transformer network is constrained by physical information to ensure that the network can learn the inherent dynamic characteristics of the thermal power unit and produce advanced prediction results of key parameters that conform to the mechanism of the unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the implementation principle of the present invention; Figure 2 Schematic diagram of the Transformer encoder for extracting long-term dependencies between key parameters and system inputs; Figure 3 Schematic diagram of the Transformer decoder for multi-step prediction of key parameters; Figure 4 Schematic diagram of multi-step prediction model training based on discrete mechanism model regularization. DETAILED DESCRIPTION

[0019] The present invention will be further described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0020] like Figure 2 The principle diagram of the modeling method for advanced prediction of key parameters of thermal power units based on physical information Transformer is shown. The present invention provides a method for advanced prediction of key parameters of coal-fired thermal power units based on physical information machine learning and Transformer neural network, which includes a Transformer encoder for extracting long-term dependency between key parameters and system input, a Transformer decoder for multi-step prediction of key parameters, and a multi-step prediction loss function based on regularization of discrete mechanism model, which can be divided into two implementation stages: offline training and online prediction.

[0021] like Figure 1 As shown, a method for advanced prediction of key parameters of thermal power units based on physical information Transformer includes: S1: Obtain the historical operation data of the unit and divide the historical operation data of the unit into labeled time series samples.

[0022] against Figure 2 The offline training phase is shown in Figure 1. First, the historical operation data of the unit is obtained from the unit DCS control system, and the historical operation data of the unit is divided into labeled time series samples. For the historical data of the unit with a time length of T and a sampling period of Δt, each time step , A training sample can be extracted, where p is the length of historical data in each sample and f is the number of forward prediction steps. t -p to t The unit load, drum pressure, main steam pressure, coal supply, and valve opening of the step are taken as a time series sample X t =[x t-p ,x t-p+1 ,…,x t ]. t +1 to t +f step unit load, drum pressure, main steam pressure as the corresponding label Y t , t arrivet +f-1 step unit load, drum pressure, main steam pressure, coal supply, and regulating valve opening are used as auxiliary variables A for calculating physical information constraints. t , build a labeled training dataset .

[0023] S2: Encode the time series samples into features containing the long-term historical status and input information of the unit for representation.

[0024] There will be labeled training samples X t Input Transformer encoder, extract long-distance dependencies to form feature representation z containing the long-term historical state of the unit and input information N .

[0025] (25) (26) (27) Among them, SA stands for self-attention mechanism, MLP stands for multi-layer perceptron, LN stands for layer normalization operation, E is the weight matrix of the embedding layer, and E pos is the position encoding, and N is the number of self-attention layers.

[0026] S3: Advance prediction of key parameters of the unit; Feature representation based on long-term historical information N , the Transformer decoder is used to make advance predictions of the key parameters of the unit for the next f steps.

[0027] (28) (29) (30) (31) (32) Among them, CA represents the mutual attention mechanism, Linear represents the linear layer with learnable parameters, and M is the number of mutual attention layers. It is the multi-step prediction value of the key parameters of the unit.

[0028] S4: Based on the mechanism model of thermal power units, a physical information regularization term is constructed to obtain a loss function with physical information regularization.

[0029] Get advanced prediction results After that, the unit mechanism model is discretized by the Euler method, the physical information regularization term is constructed, and the loss function with physical information regularization is calculated. The thermal power unit mechanism model used is: (33) (34) (35) Among them, k2 and k3 are normalization coefficients, t e It is the time constant of the turbine valve adjustment process. T is the total volume of the working fluid in each stage of superheater, ρ T is the steam density in the superheater, C0 is the damping coefficient in the superheater, ρ D is the density of saturated steam accumulated in the drum. V s is the volume of the gas phase in the drum, Q net,ar The low calorific value of the coal entering the furnace. h s is the specific enthalpy of saturated steam, h fw is the specific enthalpy of the feed water at the evaporation system inlet, and q4 is the heat loss on the boiler side.

[0030] The continuous partial differential equations (33)(34)(35) are discretized to a fixed time step Δt using the Euler method, and the predicted values ​​are , the truth value Y t and auxiliary variable A t Substitute it into the discretized equation and construct it into a physical information regularization term through the L2 norm.

[0031] (36) (37) (38) in, , , yes The advanced prediction values ​​of medium load, drum pressure and main steam pressure, and Yes A t Medium coal feeding amount and regulating valve opening value, , and Yes t True values ​​of medium load, drum pressure and main steam pressure.

[0032] Through the regularization term Py p , Py pD , Py pT Constrained Forecast Sequence and the true value Yt The mean squared error loss function is: (39) Among them, S is the total number of training samples.

[0033] S5: The Transformer model for advanced prediction of key parameters of thermal power units is trained through the mean square error loss function with physical information regularization term.

[0034] The Transformer model for advanced prediction of key parameters of thermal power units includes: An encoder, the encoder comprising an input embedding layer and N stacked self-attention layers, the encoder takes historical time series data of operating parameters such as unit load, drum pressure, main steam pressure, coal feed rate, and throttle opening as input, extracts long-distance dependencies through the embedding layer and the self-attention layer, and forms a feature representation containing the long-term historical state of the unit and input information; The decoder includes an input embedding layer, M stacked mutual attention layers and an output linear transformation layer. The decoder takes the historical time series of the key parameters to be predicted and the future time series initialized to 0 as input, combines the feature representation output by the encoder, and predicts the future time series through the embedding layer, the mutual attention layer and the output layer.

[0035] S51: Set the training set D={X t ,Y t ,A t} training sample X t Input the Transformer advance prediction model to obtain the advance prediction results of the key parameters of the unit; S52: Calculate the advance prediction result and label Y t The mean square error loss between the forward prediction results and the auxiliary variable A t and system parameters are used to calculate the value of the physical information regularization term; Multiply the physical information regularization term by the weight λ and add it to the mean square error true value loss to get the total loss function value; S53: Input the total loss function value into the Adam optimizer to update the parameters of the Transformer model; S54: Repeat steps S51-S53 until the total loss function value basically stops decreasing, and finally obtain the Transformer advance prediction model with the best prediction effect.

[0036] The Transformer model for advanced prediction of key parameters of thermal power units is trained through a mean square error loss function with a physical information regularization term, so that the Transformer model can learn the dynamic characteristics of parameter changes, produce prediction results that are consistent with the unit's operating mechanism, and improve the prediction accuracy and generalization of the model.

[0037] Finally, the deep learning optimizer is used to update the Transformer model parameters according to the loss function value. Samples are repeatedly input into the Transformer model and the parameters are updated until the loss function value basically stops decreasing.

[0038] against Figure 1 In the online prediction stage, the Transformer advance prediction model trained in the offline stage is deployed to the computer at the unit site. The online real-time data of the unit is obtained through the DCS system, including the unit load, drum pressure, main steam pressure, coal feed rate, and regulating valve opening in the past p steps. By inputting it into the Transformer model, the advance prediction results of the key parameters of the unit can be obtained.

[0039] like Figure 2 The figure shows a schematic diagram of the Transformer encoder for extracting long-term dependencies between key parameters and system inputs. After the unit's historical operation data sample Xt is input into the encoder, it first passes through the embedding layer to form the encoder input embedding and is added to the position code. Then, N self-attention layers are used to extract the long-term dependency feature representation of the unit's historical information. Each self-attention layer contains two modules: self-attention and multi-layer perceptron. The self-attention module consists of three parts: residual connection, layer normalization, and self-attention calculation. The multi-layer perceptron module consists of three parts: residual connection, layer normalization, and multi-layer perceptron.

[0040] like Figure 3 The following is a schematic diagram of the key parameter multi-step prediction Transformer decoder. The decoder input is a vector sequence of length l+f, and the first l vectors are t -l to tThe label value of the decoder, and the last f vectors are the key parameter values ​​to be predicted, which are initialized to zero vectors. The decoder input is mapped to the decoder input embedding through the embedding layer and added to the position code. Then, the feature representation of the target future key parameter sequence is decoded through M mutual attention layers, and finally the advance prediction result is obtained through layer normalization and learnable linear transformation. The mutual attention layer consists of a self-attention module, a mutual attention module, and a multi-layer perceptron module. The self-attention module consists of three parts: residual connection, layer normalization, and self-attention calculation, which extracts the self-features of the decoder input embedding. The mutual attention module consists of three parts: residual connection, layer normalization, and self-attention calculation, which is used to model the correlation between the decoder input embedding and the long-term dependency feature representation of the unit history information output by the encoder. The multi-layer perceptron module consists of three parts: residual connection, layer normalization, and multi-layer perceptron, which further maps the correlation between the unit history information captured by mutual attention and the target future sequence into the feature representation of the target future sequence.

[0041] S6: Obtain the online real-time data of the unit, input the trained Transformer model for advanced prediction of key parameters of the thermal power unit, and obtain the advanced prediction results of the key parameters of the unit.

[0042] S61: Deploy the Transformer model for predicting the key parameters of the thermal power unit trained in the offline stage to the computer on the unit site, and obtain the online real-time data of the unit through the DCS system, including the unit load, drum pressure, main steam pressure, coal supply, and regulating valve opening in the past p steps; S62: The unit load, drum pressure, main steam pressure, coal feed rate, and regulating valve opening of the past p steps are input into the Transformer model for advanced prediction of key parameters of the thermal power unit to obtain the advanced prediction results of the key parameters of the unit.

[0043] like Figure 4 The figure shows a schematic diagram of multi-step prediction model training based on discrete mechanism model regularization. The training data set is a set of triples D = {X t ,Y t ,A t}. Training sample X t Input the Transformer advance prediction model to obtain the advance prediction results of the key parameters of the unit. The advance prediction results are consistent with the label Y t The mean square error truth loss is calculated to evaluate the prediction accuracy in a purely data-driven way. On the other hand, the advance prediction results are compared with the auxiliary variable A tThe system parameters are used to calculate the value of the physical information regularization term, and evaluate whether the prediction results are consistent with the system dynamic characteristics from the perspective of the system mechanism, thereby forcing the model to learn the dynamic characteristics of the system. The physical information regularization term multiplied by the weight λ is added to the mean square error true value loss to obtain the loss function value, which is input into the Adam optimizer to update the parameters of the Transformer model to obtain the optimal Transformer advance prediction model.

[0044] In summary, the present invention proposes a method for advanced prediction of key parameters of coal-fired thermal power units based on physical information machine learning and Transformer neural network. The Transformer encoder based on the self-attention mechanism extracts the long-term dependency of key parameters such as coal-fired unit load, drum pressure, main steam pressure, and historical data of control inputs such as coal feed rate and throttle opening to form a characteristic representation of the unit state; based on the long-term characteristic representation of the unit state, the Transformer decoder based on the mutual attention mechanism is used to predict key parameters such as unit load, drum pressure, main steam pressure, etc. in the future multiple steps; the mechanism model of the unit is discretized by the Euler method to construct a physical information regularization term, and the training of the Transformer network is constrained by physical information to ensure that the network can learn the inherent dynamic characteristics of the thermal power unit and produce advanced prediction results of key parameters that conform to the mechanism of the unit.

Claims

1. A method for advanced prediction of key parameters of thermal power units based on physical information Transformer, characterized in that: include: S1: Obtain the historical operation data of the unit and divide the historical operation data of the unit into labeled time series samples; S2: Encode the time series samples into features containing the long-term historical state and input information of the unit for representation; S3: Advance prediction of key parameters of the unit; S4: Based on the mechanism model of thermal power units, a physical information regularization term is constructed to obtain a loss function with physical information regularization; S5: Train the Transformer model for advanced prediction of key parameters of thermal power units through the mean square error loss function with physical information regularization term; S6: Obtain the online real-time data of the unit, input the trained Transformer model for advanced prediction of key parameters of the thermal power unit, and obtain the advanced prediction results of the key parameters of the unit.

2. The method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 1 is characterized in that: In step S1, for the historical data of the unit with a time length of T and a sampling period of Δt, each time step A training sample can be extracted, where p is the length of historical data in each sample and f is the number of forward prediction steps; by t -p to t The unit load, drum pressure, main steam pressure, coal supply, and valve opening of the step are taken as a time series sample X t =[x t-p ,x t-p+1 ,…,x t ]; t +1 to t +f step unit load, drum pressure, main steam pressure as the corresponding label Y t , t arrive t +f-1 step unit load, drum pressure, main steam pressure, coal supply, and regulating valve opening are used as auxiliary variables A for calculating physical information constraints. t , build a labeled training dataset .

3. The method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 1 is characterized in that: The step S2 comprises: There will be labeled training samples X t Input Transformer encoder, extract long-distance dependencies to form feature representation z containing the long-term historical state of the unit and input information N .

4. The method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 3 is characterized in that: The step S3 comprises: The feature representation of long-term historical information is z N The input Transformer decoder makes advance predictions on the key parameters of the unit for the next f steps.

5. The method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 1 is characterized in that: In step S4, the mechanism model of the thermal power unit is: Among them, k2 and k3 are normalization coefficients, t e It is the time constant of the turbine valve adjustment process. T is the total volume of the working fluid in each stage of superheater, ρ T is the steam density in the superheater, C0 is the damping coefficient in the superheater, ρ D is the density of saturated steam accumulated in the drum, V s is the volume of the gas phase in the drum, Q net,ar The low calorific value of the coal entering the furnace. h s is the specific enthalpy of saturated steam, h fw is the specific enthalpy of feed water at the evaporation system inlet, and q4 is the heat loss on the boiler side.

6. The method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 1 is characterized in that: In step S4, the loss function is: in, To constrain the prediction sequence, Y t is the true value, S is the total number of training samples; , , yes The advanced prediction values ​​of medium load, drum pressure and main steam pressure, and Yes A t Medium coal feeding amount and regulating valve opening value, , and Yes t The true values ​​of medium load, drum pressure and main steam pressure, k2 and k3 are normalization coefficients, t e It is the time constant of the turbine valve adjustment process. T is the total volume of the working fluid in each stage of superheater, ρ T is the steam density in the superheater, C0 is the damping coefficient in the superheater, ρ D is the density of saturated steam accumulated in the drum, V s is the volume of the gas phase in the drum, Q net,ar The low calorific value of the coal entering the furnace. h s is the specific enthalpy of saturated steam, h fw is the specific enthalpy of feed water at the evaporation system inlet, and q4 is the heat loss on the boiler side.

7. The method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 1 is characterized in that: The Transformer model for advanced prediction of key parameters of thermal power units includes: An encoder, the encoder comprising an input embedding layer and N stacked self-attention layers, the historical time series data of the encoder operation parameters being used as input, the long-distance dependencies being extracted through the embedding layer and the self-attention layer, and forming a feature representation containing the long-term historical state of the unit and the input information; The decoder includes an input embedding layer, M stacked mutual attention layers and an output linear transformation layer. The decoder takes the historical time series of the key parameters to be predicted and the future time series initialized to 0 as input, combines the feature representation output by the encoder, and predicts the future time series through the embedding layer, the mutual attention layer and the output layer.

8. According to the method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 7, the step S5 comprises: S51: Set the training set D={X t , Y t , A t } training sample X t Input the Transformer advance prediction model to obtain the advance prediction results of the key parameters of the unit; S52: Calculate the advance prediction result and label Y t The mean square error loss between the forward prediction results and the auxiliary variable A t and system parameters are used to calculate the value of the physical information regularization term; Multiply the physical information regularization term by the weight λ and add it to the mean square error true value loss to get the total loss function value; S53: Input the total loss function value into the Adam optimizer to update the parameters of the Transformer model; S54: Repeat steps S51-S53 until the total loss function value basically stops decreasing, and finally obtain the Transformer advance prediction model with the best prediction effect.

9. The method for advanced prediction of key parameters of thermal power units based on physical information Transformer according to claim 1 is characterized in that: The step S6 comprises: S61: Deploy the Transformer model for predicting the key parameters of the thermal power unit trained in the offline stage to the computer on the unit site, and obtain the online real-time data of the unit through the DCS system, including the unit load, drum pressure, main steam pressure, coal supply, and regulating valve opening in the past p steps; S62: The unit load, drum pressure, main steam pressure, coal feed rate, and regulating valve opening of the past p steps are input into the Transformer model for advanced prediction of key parameters of the thermal power unit to obtain the advanced prediction results of the key parameters of the unit.

Images