CN-120955621-B - Electric power and electricity balance calculation method and system based on machine learning

CN120955621BCN 120955621 BCN120955621 BCN 120955621BCN-120955621-B

Abstract

The invention discloses a power and electricity balance calculation method and a system based on machine learning, which belong to the technical field of power system optimization scheduling, a deep neural network is combined with optimization calculation, the running state of most thermal power units is determined in advance through a Seq2Seq-Attention model by adopting an integer variable reduction strategy, the input of the model is relaxation solution of the running state and scene curve data, and the output is the running state of the thermal power units, so that the effect of improving the calculation speed of the power and electricity balance problem is achieved, the accelerating effect of a rapid calculation method is verified through a designed accelerating effect evaluation index and an accelerating effect verification method based on variable time scale scene generation, the accelerating effect is high in accuracy, and the practicability of the method is high.

Inventors

WANG JIANXUE
ZHU ZIYI
HUANG CHENYANG
Gong Leteng
LI XINGYI
ZHANG QI
LI HAOTIAN

Assignees

西安交通大学

Dates

Publication Date: 20260508
Application Date: 20250730

Claims (6)

1. The electric power and electricity balance calculation method based on machine learning is characterized by comprising the following steps of: The thermal power generating unit start-stop plan is used as a sample label, and the deep neural network is trained by combining scene curve data and running state relaxation solution, wherein the scene curve data comprises a load curve, a connecting line curve, a reservoir capacity type hydroelectric power curve, a radial flow type hydroelectric resource curve, a wind power resource curve and a photovoltaic resource curve; The method comprises the steps of inputting curve data of a scene to be predicted and relaxation solutions of running states of a thermal power generating unit into a trained deep neural network, outputting the predicted running states of the thermal power generating unit through a thermal power generating unit state analysis model, and outputting predicted starting probability of the thermal power generating unit through a Seq2Seq-Attention model by the thermal power generating unit state analysis model And then will Mapping to predicted operating conditions The mapping rule is: Wherein, the A confidence threshold value is predicted for the thermal power generating unit, The range of the values is as follows ; Based on the obtained predicted operating state Checking and adjusting the operation state by a state post-adjustment method in combination with the shortest start-stop time constraint of the thermal power generating unit to obtain an adjusted operation state meeting the constraint, wherein the state post-adjustment method comprises the following steps: predicted running state of thermal power generating unit Aggregation is carried out according to the time period, and the time points which are the same in the predicted running state and adjacent to each other are aggregated into one time period to obtain And , wherein, The predicted running state of the thermal power generating unit i in the s-th time period is represented, and the value range is , Representing the length of the s-th time period; Calculating the time length of a continuous time period with the predicted running state of-1 before and after the current time period; Traversing the period of time obtained by polymerization, if Directly skipping the current time period if The length of the current time period plus the calculated time length meets the constraint of the shortest start-up and stop time, and the current time period does not need to be adjusted; And fixing the obtained adjusted running state, reducing integer variables in the mathematical optimization model, and obtaining a final machine set start-stop plan and a final machine set output scheme through optimization solution.
2. The machine learning-based power and quantity balance calculation method according to claim 1, wherein the thermal power generating unit state analysis model comprises: The starting probability analysis model adopts a Seq2Seq-Attention model, takes an LSTM network as the basis of an encoder-decoder, introduces an additive Attention mechanism, inputs the relaxation solution of the running state and scene curve data, and outputs the predicted starting probability of the thermal power generating unit ; Running state mapping model, predicting starting probability Mapping to predicted running state of thermal power generating unit 。
3. The machine learning-based power and electricity balance calculation method according to claim 2, wherein a power-on probability analysis model is adopted to obtain a predicted power-on probability of the thermal power generating unit , The starting probability of the thermal power unit i at the time t is represented, and the value range is the decimal between [0,1 ]; encoding, by the encoder, the relaxed solution of the run state and the scene curve data into a fixed length context vector; And then the context vector is decoded by the decoder to obtain the predicted startup probability of the thermal power generating unit, and a attention mechanism is introduced in the encoding and decoding processes, so that the decoder is allowed to consider information of different positions in the output of the encoder when generating the output of each moment.
4. The machine learning based power balance calculation method of claim 2 wherein in the operational state analysis model, the interpretability of the model is enhanced by: in the startup probability analysis model, the input of the model comprises a relaxation solution of the running state and scene curve data; the weights of different positions are considered to be input through an Attention mechanism, and a relaxation solution of an operation state is adjusted through a deep learning model; in the operation state mapping model, the operation state of the partial thermal power generating unit is predicted.
5. The machine learning-based electric power and electricity balance calculation method according to claim 1, wherein the final machine set start-stop plan and the output plan are obtained by optimizing and solving, and the acceleration effect is verified based on a variable time scale scene generation method, specifically comprising the following steps: Generating daily average power based on a variation self-encoder, sampling a variable time scale scene by a Markov chain, and generating daily fluctuation power by a vector autoregressive model; the verification indexes comprise running state identification rate, prediction accuracy rate, acceleration rate, objective function error and balance margin error.
6. A machine learning-based power balance computing system, comprising: The scene module is used for solving a historical scene through the electric power and electric quantity balance model to generate a unit start-stop plan and a power output scheme, taking the thermal power unit start-stop plan as a sample label, and combining scene curve data and an operation state relaxation solution training deep neural network, wherein the scene curve data comprises a load curve, a tie line curve, a reservoir capacity type hydroelectric power curve, a radial flow type hydroelectric resource curve, a wind power resource curve and a photovoltaic resource curve; The training module inputs curve data of a scene to be predicted and the running state relaxation solution of the thermal power generating unit into the trained deep neural network, and outputs the predicted running state of the thermal power generating unit through a thermal power generating unit state analysis model, wherein the thermal power generating unit state analysis model outputs the predicted starting probability of the thermal power generating unit through a Seq2Seq-Attention model And then will Mapping to predicted operating conditions The mapping rule is: Wherein, the A confidence threshold value is predicted for the thermal power generating unit, The range of the values is as follows ; The adjustment module is based on the obtained predicted running state Checking and adjusting the operation state by a state post-adjustment method in combination with the shortest start-stop time constraint of the thermal power generating unit to obtain an adjusted operation state meeting the constraint, wherein the state post-adjustment method comprises the following steps: predicted running state of thermal power generating unit Aggregation is carried out according to the time period, and the time points which are the same in the predicted running state and adjacent to each other are aggregated into one time period to obtain And , wherein, The predicted running state of the thermal power generating unit i in the s-th time period is represented, and the value range is , Representing the length of the s-th time period; Calculating the time length of a continuous time period with the predicted running state of-1 before and after the current time period; Traversing the period of time obtained by polymerization, if Directly skipping the current time period if The length of the current time period plus the calculated time length meets the constraint of the shortest start-up and stop time, and the current time period does not need to be adjusted; and the output module is used for fixing the obtained adjusted running state, reducing integer variables in the mathematical optimization model and obtaining a final machine set start-stop plan and a final machine set output scheme through optimization solution.

Description

Electric power and electricity balance calculation method and system based on machine learning Technical Field The invention belongs to the technical field of power system optimal scheduling, and particularly relates to a power and electricity balance calculation method and system based on machine learning. Background As the scale of the electric power system becomes larger and the calculation time scale is increased, the calculation difficulty of the electric power and electric quantity balance problem is remarkably increased, the safety and the economical efficiency of the operation of the electric power system are comprehensively considered by a traditional mathematical optimization method, and the electric power and electric quantity balance model is solved to obtain an on-off plan and an output scheme of each type of unit, but the on-off plan and the output scheme are difficult to promote the solving speed while ensuring the solving precision. As models become finer, the number of integer variables increases, and there are cases of superposition and coupling between multiple integer variables. Therefore, the running state of the partial thermal power generating unit is determined in advance, and it becomes important to cut down integer variables in a mathematical optimization model. In a system with the new energy installation ratio of nearly 50%, the net load peak Gu Chalv is up to 86.49%, and the thermal power generating unit start-stop state variable coupling causes the time consumption of Mixed Integer Linear Programming (MILP) solving to be increased sharply (extreme scenes exceed 500 seconds). Traditional optimization methods (such as Gurobi solvers) have difficulty meeting the timeliness requirements of rolling schedule in the day. Although the existing prediction schemes such as CNN/converter increase the speed, the prediction states violate the shortest start-stop constraint because of ignoring the optimization boundary information contained in the running state relaxation solution and no physical constraint post-treatment is introduced, and the objective function error is up to 194.37%, so that the load shedding risk is caused. The existing electric power and electric quantity balance problem faces two major bottlenecks: The calculation complexity is high, as the installed ratio of the new energy source is increased (nearly 50%), the net load peak Gu Chalv is up to 86.49%, and the coupling of integer variables (the start-stop state of the thermal power unit) causes the time consumption of the mixed integer planning solution to be greatly increased (the extreme scene exceeds 500 seconds). The accuracy-speed contradiction is that although the traditional machine learning acceleration scheme (such as CNN, transformer) increases the speed, the optimization information contained in the running state relaxation solution is ignored, and the unit constraint post-processing is not introduced, so that the predicted state violates the actual running rule (such as the shortest start-stop time), and the objective function error reaches 194.37%. Disclosure of Invention The invention aims to solve the technical problems of the prior art, and provides a power and electricity balance calculation method and a system based on machine learning, which are used for solving the technical problems that the solving precision and the solving speed are difficult to be compatible in the power and electricity balance problem. The invention adopts the following technical scheme: a power and electricity balance calculation method based on machine learning comprises the following steps: The thermal power unit start-stop plan is used as a sample label, and the deep neural network is trained by combining scene curve data and running state relaxation solutions; Inputting curve data of a scene to be predicted and the running state relaxation solution of the thermal power unit into the trained deep neural network, and outputting the predicted running state of the thermal power unit through a thermal power unit state analysis model; based on the obtained predicted running state, the running state is checked and adjusted by a post-state adjustment method in combination with the running constraint of the thermal power unit, and the adjusted running state meeting the constraint is obtained; And fixing the obtained adjusted running state, reducing integer variables in the mathematical optimization model, and obtaining a final machine set start-stop plan and a final machine set output scheme through optimization solution. Preferably, the scene curve data includes: load curve, tie line curve, reservoir capacity type hydroelectric power curve, radial flow type hydroelectric resource curve, wind power resource curve and photovoltaic resource curve. Preferably, the thermal power generating unit state analysis model includes: The starting probability analysis model adopts a Seq2Seq-Attention model, takes an LSTM network as the basis of an