Search

CN-122026525-A - Prediction decision integrated resource scheduling method and system for electric-hydrogen-heat integrated energy system

CN122026525ACN 122026525 ACN122026525 ACN 122026525ACN-122026525-A

Abstract

The invention discloses a prediction decision integrated resource scheduling method and a system for an electric-hydrogen-heat integrated energy system, comprising the following steps of establishing an electric-hydrogen-heat integrated energy system resource scheduling optimization model; the method comprises the steps of constructing a prediction network of uncertain parameters required by a resource scheduling optimization model, predicting future photovoltaic output and load demands by utilizing historical multi-day data in a sliding window mode, respectively inputting the prediction data and the real data into a solver to solve a scheduling problem, constructing a decision guiding loss function by comparing operation cost differences of the prediction data and the real data, calculating gradients of the loss function on the prediction network parameters by adopting a random disturbance gradient estimation method, iteratively updating the network parameters by utilizing a back propagation algorithm, predicting future scenes by utilizing the trained prediction network, and inputting a prediction result into the solver to obtain an optimal scheduling scheme and minimum operation cost. Compared with the traditional method, the method can reduce the daily operation cost by 0.90% -6.30%.

Inventors

  • YU LIANG
  • ZHU RUI
  • CHEN ZHIQIANG
  • YUE DONG

Assignees

  • 南京邮电大学

Dates

Publication Date
20260512
Application Date
20260410

Claims (8)

  1. 1. The prediction decision integrated resource scheduling method for the electric-hydrogen-heat comprehensive energy system is characterized by comprising the following steps of: Step 1, establishing an electricity-hydrogen-heat comprehensive energy system resource scheduling optimization model which comprises an operation cost objective function, decision variables and constraint conditions; step 2, constructing a prediction network of uncertain parameters required by the electricity-hydrogen-heat comprehensive energy system resource scheduling optimization model in the step 1, and predicting future photovoltaic output and load demand data by utilizing historical multi-day data in a sliding window mode; Step 3, respectively inputting future photovoltaic output and load demand data predicted by the prediction network in the step 2 and real data into a solver to obtain a minimum value of an objective function of the resource scheduling problem of the electric-hydrogen-heat comprehensive energy system, and calculating an error between the minimum value and the minimum value to construct a loss function; Step 4, calculating the gradient of the error in the step 3 on the predicted network parameters by adopting a gradient estimation method, and updating the predicted network parameters by utilizing a back propagation algorithm; and 5, predicting a future scene by utilizing the prediction network trained in the step 4, and inputting a prediction result into a solver to solve an optimal scheduling scheme and minimum running cost.
  2. 2. The prediction decision integrated resource scheduling method for the electric-hydrogen-heat integrated energy system according to claim 1, wherein, The decision variable in the step 1 Comprising Time slot buying electric power to electric network 、 Time slot selling electric power to a grid 、 Time slot hydrogen market volume 、 Input power of time slot electrolyzer 、 Output power of slotted fuel cell 、 Time slot electric boiler input power 、 Charging power of time slot energy storage battery 、 Discharge power of time slot energy storage battery 、 Time slot hot water tank charging power 、 Thermal power of hot water tank of time slot 、 Operating state of the electrolysis cell of the time slot 、 Operating state of a slotted fuel cell ; The constraint conditions include: ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; Wherein, the Representation of A hydrogen storage level of the time slot hydrogen storage tank; Representation of A hydrogen storage level of the time slot hydrogen storage tank; Indicating a maximum hydrogen storage level of the hydrogen storage tank; indicating the electro-hydrogen conversion efficiency of the electrolytic cell; Represents the hydrogen-to-electricity conversion efficiency of the fuel cell; And Respectively representing the maximum input power of the electrolytic tank and the maximum output power of the fuel cell; indicating that the electrolyzer and the fuel cell cannot operate simultaneously; representing the maximum amount of hydrogen available from the hydrogen market; Representation of Energy storage level of the time slot hot water tank; Representation of The energy storage level of the time slot heat storage tank; Representing a maximum energy storage level of the hot water tank; And Respectively representing the maximum heat filling power and the maximum heat releasing power of the hot water tank; And Respectively representing the heat filling efficiency and the heat releasing efficiency of the hot water tank; Indicating that the hot water tank can not run simultaneously with the heat release; Representation of The energy storage level of the time slot energy storage battery; And Respectively represent The charging power and the discharging power of the time slot energy storage battery; And Respectively representing the charging efficiency and the discharging efficiency of the energy storage battery; Representation of The energy storage level of the time slot energy storage battery; Representation of The energy storage level of the time slot energy storage battery; Representing a maximum energy storage level of the energy storage battery; And Respectively representing the maximum charging power and the maximum discharging power of the energy storage battery; Indicating that the charging and discharging of the energy storage battery cannot be operated simultaneously; And Respectively representing the heat production efficiency of the electric boiler and the maximum input power of the electric boiler; 、 、 And Respectively represent Buying power, selling power, photovoltaic power generation power and electrical load requirements of a time slot; And Respectively represent Thermal energy and thermal load demands generated by a slotted fuel cell; Representation of The time slot electric boiler outputs heat energy; And Respectively representing the heat recovery efficiency and the electrothermal conversion efficiency of the fuel cell; 、 And Representing the operating, start-up and shut-down cost coefficients of an electrolyzer or fuel cell in a hydrogen energy storage system, 、 And Respectively represent The operation, start-up and shutdown conditions of the slot electrolyzer or fuel cell, 、 And The hydrogen purchase price, the electricity purchase price and the electricity selling price of the t time slot are respectively represented; The objective function is: ; ; ; ; Wherein, the 、 、 Respectively represent The operation cost, the hydrogen purchasing cost and the power grid interaction cost of the time slot hydrogen energy storage system, In (a) and (b) Indicating the equipment electrolyzer and the fuel cell.
  3. 3. The prediction decision integrated resource scheduling method for the electric-hydrogen-thermal integrated energy system according to claim 1, wherein the training set of the prediction network in the step 2 is prepared by the following method: a1 Constructing an original characteristic sequence by adopting a historical data set containing three characteristics of photovoltaic output, electric load and thermal load, recording a first step The number of the characteristic channels is one, Wherein For the total number of features, in the first The original physical value of each time slot is ; A2 Maximum-minimum normalization processing, namely mapping the original characteristic sequence to the data of different energy sources to eliminate the dimensional difference of the data Interval, set up characteristic channel Maximum value on training set is Minimum value of Then (1) Normalized data of time of day The calculation formula is as follows: ; all inputs and internal operations of the prediction network are based on the normalized data Proceeding; a3 Constructing time sequence input tensor and label, namely constructing training samples by adopting sliding window method Setting the historical days as the days of multiple days Predicting future day Hour data, input tensor For each sample, intercept the channels before Normalized data of the day, forming dimensions of Is stacked as in batch training Tensor of (2), wherein Is of batch size, real label For the day immediately after the input window The hour normalization data is used as a supervision training target of the network; a4 Inverse normalization of the predicted results when the network outputs normalized predicted values Then, the final prediction decision is obtained by restoring the final prediction decision into the original physical dimension through the following formula : 。
  4. 4. The prediction decision integrated resource scheduling method for electric-hydrogen-thermal integrated energy system according to claim 3, wherein the prediction network in step 2 normalizes the input tensor As input, for each characteristic channel The treatment is independently carried out according to the following steps: b1 Sequence decomposition, processing the input sequence with a moving average decomposition module, for a channel Is input sequence of (a) Smoothing using a sliding window of fixed size to separate it into seasonal components of volatility And trend component And meet the following ; B2 Linear transformation prediction by applying a learnable linear transformation to the decomposed components, respectively, defining a seasonal component weight matrix And a trending component weight matrix The network directly maps out the future through matrix operation Normalized predicted components at each time, the calculation formula is expressed as: ; b3 Output splicing-all the network will Normalized prediction result obtained by calculation of each channel And (3) splicing, outputting a complete predicted tensor, and then restoring the predicted tensor into physical quantity according to an inverse normalization formula for resource scheduling decision of the system.
  5. 5. The prediction decision integrated resource scheduling method for the electric-hydrogen-heat integrated energy system according to claim 1, wherein the construction flow of the loss function in the step 3 is as follows: first, the network output value is predicted And the corresponding parameter true value Respectively inputting the model constructed in the step 1 into a CPLEX solver to obtain scheduling decisions based on prediction data And theoretical optimal scheduling decisions based on real data ; Then both are calculated in the same real environment The difference in running costs defines the decision loss function, namely: ; Wherein, the In order to run the cost objective function, Is a neural network parameter.
  6. 6. The prediction decision integrated resource scheduling method for the electric-hydrogen-heat integrated energy system according to claim 5, wherein the gradient estimation method in step 4 is specifically: For predicted network output values In each gradient estimation, a random disturbance vector conforming to a standard normal distribution is generated And constructing a predicted value after disturbance Wherein A small positive scalar, then Inputting the decision after disturbance to a solver Using a differential formula Calculating decision variables For predicted values Repeating the process And averaging the results to finally obtain a stable gradient estimated value 。
  7. 7. The prediction decision integrated resource scheduling method for the electric-hydrogen-heat integrated energy system according to claim 1, wherein the training method of the prediction network in the step 4 is as follows: Firstly, calculating a decision loss function based on a chain rule For predicted network output values The calculation formula is as follows: ; Wherein, the Gradient coefficients for decision variables for the target cost function pair, To estimate the gradient through random disturbance using the previous steps, then further calculate the decision loss function versus predictive network internal trainable parameters using the automatic differential characteristics of the deep learning framework Pytorch Is a total gradient of (2): ; Wherein, the Representing a forward propagation mapping function of the predictive network; Representing the gradient of the predicted network output value to the neural network parameter; calculated by Pytorch automatic differentiation mechanism, and finally, based on the calculated total gradient And (3) carrying out iterative updating on the network parameters by adopting a gradient descent algorithm: ; Wherein, the Is the learning rate.
  8. 8. An integrated resource scheduling system for prediction decision of an electric-hydrogen-thermal comprehensive energy system, which is used for realizing the scheduling method according to any one of claims 1 to 7, and is characterized in that the system comprises a prediction module, an optimization solving module and a cooperative training module; the prediction module receives historical multi-source data, and generates future time sequence predicted values of photovoltaic, electric load and thermal load by utilizing a neural network comprising a moving average decomposition layer and a linear transformation layer after normalization processing; the optimization solving module receives the predicted value or the historical true value output by the predicting module as a parameter, and invokes the solver to calculate a resource scheduling scheme which meets the constraint of the system and has the lowest running cost; the collaborative training module is responsible for calculating decision loss, estimating the gradient of the optimization solving module by using a random disturbance method, and updating the network parameters of the prediction module by using a back propagation algorithm.

Description

Prediction decision integrated resource scheduling method and system for electric-hydrogen-heat integrated energy system Technical Field The invention belongs to the field of integrated energy system optimal scheduling and artificial intelligence intersection, and particularly relates to a prediction decision integrated resource scheduling method and system for an electric-hydrogen-heat integrated energy system. Background In recent years, with the advancement of energy transformation, integrated energy systems have been rapidly developed. The electric-hydrogen-heat integrated energy system is used as a novel energy system, and the coordination optimization of various energy forms is realized by coupling the electric power system, the hydrogen energy system and the thermodynamic system, so that the energy utilization efficiency can be effectively improved, the renewable energy consumption can be promoted, and the system operation cost can be reduced. The hydrogen energy is used as a clean energy carrier, can be converted into hydrogen energy through an electrolytic tank for storage, is converted back into electric energy through a fuel cell when needed, and simultaneously generates heat energy to realize space-time transfer and multi-energy complementation of energy. Therefore, the research has important economic and social values. However, the optimal scheduling of electrical-hydrogen-thermal integrated energy systems presents a number of challenges. First, the uncertainty of both sides of the source load and the complexity of the system coupling increase the scheduling difficulty. The electro-hydro-thermal system relates to fluctuation of output of renewable energy sources such as photovoltaics, wind power and the like, and dynamic change of various load demands of electricity, heat and hydrogen. Although the traditional mathematical programming method (such as robust optimization and random optimization) can handle uncertainty, the robust optimization usually makes decisions based on severe scenes, so that the scheduling scheme is biased to be conservative, and the random optimization depends on scene generation and reduction, so that the calculated amount is large. In addition, although the simple data-driven prediction method can capture the time sequence characteristics, the method has limitations in processing the physical constraint and coupling relation between the multi-energy streams. And secondly, the prediction and decision links in the existing scheduling system are relatively independent. The current scheduling strategy mostly follows a mode of 'prediction before decision', namely, a prediction model is trained by taking a minimized statistical error as a target, and then a prediction result is used as input to carry out optimization solution. Since the loss function of the prediction model does not contain subsequent scheduling cost information, the statistical prediction error minimization is not equivalent to the target optimization of the scheduling layer. The target deviation of the prediction layer and the decision layer may cause the situation that the final scheduling scheme has cost increase or constraint violation in actual operation, and the comprehensive benefit of the electric-hydrogen-thermal system is difficult to fully develop. 1. The application provides a monthly electricity sales prediction method, a monthly electricity sales prediction device and a server, which are searched by Chinese patent with publication number CN108629625A, and the application is different from the application as follows: 1. The patent CN108629625A mainly relates to a power supply enterprise and a power data server, and focuses on a single economic index of monthly electricity sales, wherein the application is a complex optimization problem of multiple devices, multiple energy sources and multiple constraints; 2. The patent CN108629625A adopts a two-stage mode of basic prediction and correction factors, predicts a monthly electricity sales basic value by using a Holt-winter model, corrects by introducing additional factors including air temperature, random variation and the like by using a multiple linear regression model, is a series combination of two prediction models, does not relate to a decision link and does not have any feedback mechanism; 3. Patent CN108629625A mentions that the basis for decision making by the power supply enterprises is not technically coupled to any specific decision model or optimization algorithm by introducing additional factors for explicit processing. The method has the advantages that the influence factors such as air temperature, random variation, spring festival and the like are used as additional variables to be introduced into the linear regression model, the uncertainty is captured in a prediction layer, the prediction model is a trainable component of a decision system, the model is learned and used for coping with the uncertainty in decision results, the meth