Decision-oriented source load multi-layer optimization prediction method fusing nonlinear features

Abstract

A decision-oriented source load multi-layer optimization prediction method fused with nonlinear features comprises the steps that after a double-layer parameter optimization problem is constructed and initialized, hidden layer parameters obtained through calculation are input into a two-stage optimization model, and after output layer parameters are obtained through an operation cost optimal target optimization algorithm, the two-stage optimization model is optimized; and new hidden layer parameters are updated and generated through a genetic algorithm, and loop optimization is performed until a double-layer parameter combination, namely hidden layer nonlinear mapping parameters and output layer linear weight parameters, which enables the operation cost and the prediction error to reach optimal balance is obtained, so that optimal prediction is realized. According to the method, a prediction feature mapping method of a hidden layer of a reference neural network is introduced, nonlinear mapping is carried out on source load power prediction features, and the weight, bias and activation functions of mapping are optimized by taking a two-stage operation target as a guide through an evolutionary algorithm; the obtained double-layer parameter combination improves the depiction capability of the prediction model on multi-energy coupling nonlinear features through feature mapping, and achieves the dynamic balance of the park energy scheduling economy and prediction precision through two-stage operation target-oriented collaborative optimization. And coordinated optimization of the operation target prediction modeling parameters of the park integrated energy system is realized.

Description

Technical Field

[0001] The present invention relates to a technology in the field of resource allocation, specifically a decision - oriented multi - layer optimization prediction method for source - load integration with non - linear characteristics. Background Art

[0002] Existing methods are based on linear regression models. Although they can construct convex optimization problems for solution, they cannot effectively represent non - linear characteristics such as spatio - temporal coupled non - stationary correlations, asymmetric diffusion of multi - energy flow interactions, and dynamic evolution of error standard deviations in source - load power prediction. Their linear mapping relationship is difficult to depict the complex correlations between multi - energy conversion constraints and cross - scale fluctuations, resulting in limited prediction accuracy and directly affecting the optimization effect of operation strategies. Summary of the Invention

[0003] Aiming at the deficiency that the existing integrated energy system operation target prediction model in the park is difficult to fully reflect the non - linear relationship between characteristics and prediction targets and has limited prediction accuracy, the present invention proposes a decision - oriented multi - layer optimization prediction method for source - load integration with non - linear characteristics. By introducing a prediction feature mapping method for the hidden layer of the reference neural network, the source - load power prediction features are non - linearly mapped, and through an evolutionary algorithm, the weights, biases, and activation functions of the mapping are optimized with the two - stage operation target as the guide. The obtained double - layer parameter combination improves the ability of the prediction model to depict multi - energy coupling non - linear characteristics through feature mapping, and realizes the dynamic trade - off between the economy of park energy scheduling and prediction accuracy through the collaborative optimization guided by the two - stage operation target, and realizes the coordinated optimization of the parameters of the integrated energy system operation target prediction model in the park.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to a decision - oriented multi - layer optimization prediction method for source - load integration with non - linear characteristics. After constructing and initializing a double - layer parameter optimization problem, the calculated hidden - layer parameters are input into a two - stage optimization model. After obtaining the output - layer parameters through an optimal operation - cost objective optimization algorithm, the genetic algorithm is used to update and generate new hidden - layer parameters and cycle optimization until a double - layer parameter combination that achieves the optimal balance between operation cost and prediction error is obtained, that is, the non - linear mapping parameters of the hidden layer and the linear weight parameters of the output layer, to achieve optimal prediction.

[0006] The double - layer parameter optimization problem refers to the hierarchical optimization of the parameters from the input layer to the hidden layer and from the hidden layer to the output layer. The former is globally optimized through the genetic algorithm, and the latter is dynamically adjusted by the solver in combination with a two - stage (day - ahead - intra - day optimization) operation strategy, and the two are alternately iterated to achieve global optimality.

[0007] The two-stage optimization model mentioned above refers to a collaborative optimization framework that embeds day-ahead and intra-day two-stage operation decisions, and synchronously adjusts prediction parameters and operation variables (such as energy allocation, equipment start-stop) during the parameter optimization from the hidden layer to the output layer to minimize the total operation cost.

[0008] The cost-optimal objective optimization algorithm mentioned above refers to a global search strategy based on the genetic algorithm, using the operation cost of the test set as the fitness function, optimizing the input layer parameters through the evolutionary mechanism, and combining with a two-stage solver to optimize the output layer parameters, ultimately achieving the minimization of the total operation cost.

[0009] The present invention relates to a system for implementing the above method, including: a feature mapping and non-linear transformation unit, a genetic algorithm parameter optimization unit, a two-stage operation decision unit, and a multi-objective balance regulation unit, where: the feature mapping and non-linear transformation unit performs non-linear feature mapping processing through an activation function according to the information of the original features of source and load power (such as wind speed, wind direction, time series data of electrical / thermal load), and obtains a hidden layer feature matrix; the genetic algorithm parameter optimization unit performs global optimization of the weights and biases of the input layer - hidden layer according to the hidden layer feature matrix and the feedback information of the operation cost of the test set through a genetic algorithm with a population size of 50, two-point crossover, and Gaussian mutation, and obtains the optimal input layer parameters; the two-stage operation decision unit performs collaborative optimization of the hidden layer - output layer parameters and operation variables according to the hidden layer feature matrix and the day-ahead - intra-day energy demand information through the embedded two-stage optimization model, and obtains the prediction parameters and scheduling plan for minimizing the total operation cost; the multi-objective balance regulation unit performs weight allocation processing of the operation cost and prediction error by dynamically adjusting the error weight parameter λ in the objective function according to the economic and accuracy requirement information set by the user, and obtains an optimized parameter combination suitable for different scenarios, realizing the flexible trade-off between the economy and prediction accuracy of the system. Technical effects

[0010] The present invention strengthens the mapping from features to operation cost. By mapping the features corresponding to the prediction and solving the corresponding multi-layer parameters, source and load power prediction parameters with better fitting ability for the input-output relationship can be obtained. The prediction parameters after feature mapping and multi-layer parameter optimization can not only further reduce the total operation cost of the test set, but also the prediction errors of each item of source and load power corresponding to them are reduced compared with the single-layer parameters. And by adding other terms, such as the total amount of renewable energy consumption, to the objective function, other objectives can be optimized while reducing the operation cost. Compared with the prior art, the present invention breaks through the limitations of traditional linear regression, can effectively characterize complex non-linear features such as spatio-temporal coupling fluctuations, multi-energy flow interaction asymmetric diffusion, and dynamic evolution of error standard deviation in source and load power prediction, and solves the problems that traditional linear models cannot characterize spatio-temporal coupling non-linearity, multi-energy flow interaction non-linearity, and dynamic evolution of prediction errors. Description of the drawings

[0011] Figure 1 Flowchart of the present invention;

[0012] Figure 2 It is the flow chart of genetic algorithm;

[0013] Figure 3 Flowchart for parameter optimization of double-layer source-load prediction considering hidden layer feature mapping;

[0014] Figure 4 The optimal operating cost of each generation changes with the number of evolution generations;

[0015] Figure 5 To provide the target-oriented source-load power feature mapping and the predicted effect diagram after multi-layer parameter optimization;

[0016] Figure 6 Adding different error sizes to the objective function shows the impact of the operating cost and source load prediction error. DETAILED DESCRIPTION

[0017] like Figure 1 As shown in FIG. 1 , a decision-oriented source-load multi-layer optimization prediction method integrating nonlinear features involved in this embodiment includes:

[0018] Step 1: Map the nonlinear characteristics of the source and load of the park's integrated energy system. Use the nonlinear characteristics to model and optimize the park's integrated energy system, specifically including:

[0019] 1.1 Nonlinear transformation of prediction characteristics: Wind power, photovoltaic power, electric load and thermal load in the park's integrated energy system all have their own prediction characteristics. By optimizing the prediction parameters, the linear model can be approximated to the goal of minimizing prediction accuracy or operating cost. Specifically, Y n =θ n X n ,n∈{WT,PV,Eload,Hload}, where: Y n and θ n is a vector representing the source load prediction results and prediction parameters on the dataset, with a dimension of N×1, where N is the number of samples, and X n For the matrix of predicted features, input feature matrix, dimension is N*T*k n , T is the time scale, k n is the number of features (such as wind power including wind speed, wind direction and other features); the feature parameter Y is obtained by first performing nonlinear transformation on the predicted features and then solving them n =θ' n ·H n =θ' n ·g(w n X n +b n), n ∈ {WT, PV, Eload, Hload}, H n is the hidden layer feature matrix after feature transformation, and θ' n are the parameters of the transformed features, and g(·) is the Sigmoid or ReLU function.

[0020] Through the transformation of the hidden layer, the original features of the source-load prediction are not only given new weights and biases, but also multiplied by a non-linear activation function afterwards, enhancing its ability to fit features.

[0021] 1.2 As Figure 2 shown, the input layer parameter optimization based on the genetic algorithm specifically includes:

[0022] i) For the source-load power of the park, its input feature is X n , n ∈ {WT, PV, Eload, Hload}. In order to use a matrix for the weights and biases corresponding to the features at the same time, a column of all 1 values is added to the end of the input feature: X' n = [X n 1]. In order to optimize the source-load features simultaneously, let the input layer X' = [X' WT , X' PV , X' Eload , X' Hload , that is, a matrix of size N*T*(k WT + k PV + k Eload + k Hload + 4) is established to store the weights and bias values of the features of the source-load power prediction after mapping, where: k WT , k PV , k Eload , k Hload are the corresponding feature numbers of WT, PV, Eload, and Hload respectively,

[0023] ii) Convert each of the k all features and the biases corresponding to each moment into L features on the hidden layer, and obtain the matrix H = g(AX') of the hidden layer after random feature mapping, where: g(·) is a non-linear activation function, and A is a matrix of size (k all + 4)*L. The total number of features k all = k all = k WT + k PV + k Eload + k Hload .

[0024] iii) Initialize the population: Set the population size to 50. For each type of source-load characteristics to be mapped, construct a vector of length (k + 1)*L, assign the elements inside to real numbers between [-1, 1], and then perform selection operation, crossover operation, and mutation operation in sequence.

[0025] The described selection operation, namely tournament selection, sets the tournament scale to 3. When performing the tournament operation, first, for each selection process, randomly draw 3 individuals from the population; then, compare the fitness of these 3 individuals; finally, select the individual with the highest fitness among these 3 individuals and put it into the next-generation population, and repeat this process until the number of individuals in the new population reaches the population size of 50.

[0026] iv) For the individuals that have gone through step iii), calculate their population fitness. In general optimization problems, fitness is the objective function, but for the double-layer parameter optimization problem, the selected individuals need to be transferred back to the matrix A from the input layer to the hidden layer, and the corresponding hidden layer matrix is H. The parameters from the output layer to the hidden layer are first solved for the weight matrix β according to the theory of generalized inverse matrix as the initial value of the convex optimization that embeds the two-stage operation objective. Then, call the solver to optimize the operation problem containing β and the day-ahead and intra-day two-stage decision variables to obtain the better parameter β' and the corresponding operating cost C of the test set.

[0027] v) Set the number of generations of evolution to 40, use the operating cost C of the test set as the fitness function, and repeat the crossover, mutation, and selection operations with the goal of minimizing the fitness function. After multiple generations of evolution, the individuals in the population approach the optimal solution.

[0028] The described crossover operation, namely the two-point crossover method, has a crossover probability of 0.5.

[0029] The described mutation operation, namely the Gaussian mutation method that perturbs the individual genes based on the normal distribution of the Gaussian distribution, randomly changes the genes of the individual. Assume that the individual X to be mutated in the population = [x1, x2, …, x n , where: x i (i = 1, 2, … n) represents the gene. Then, for each gene x i , a random number Δx 2 that follows the Gaussian distribution N(0, σ i ) will be generated. This random number is the perturbation amount of the mutation, and the result of the Gaussian mutation is:

[0030] Step 2, as Figure 3 shown, the multi-layer prediction parameter step-by-step optimization from the input layer to the hidden layer and from the hidden layer to the output layer specifically includes:

[0031] 2.1 Multi - layer prediction parameter optimization framework: In the source - load power prediction oriented by operating cost, the parameters from the input layer to the hidden layer and from the hidden layer to the output layer need to be solved separately. This is because in the parameter optimization of source - load power prediction with an embedded day - ahead and intra - day two - stage operation strategy, the two - stage dual variables need to be considered, and the gradient from the final operating cost to the feature parameters cannot be explicitly solved. Instead, the solver directly solves the parameters from the hidden layer to the output layer according to the optimization model with the embedded two - stage operation strategy. For the parameters from the input layer to the hidden layer, the genetic algorithm is used to select the optimal parameters. Each time the parameters from the input layer to the hidden layer are given, the solver solves the parameter optimization problem from the hidden layer to the output layer with the embedded two - stage operation strategy to obtain the corresponding operating cost under the optimal parameters; then, with the goal of minimizing the operating cost, the genetic algorithm is used to optimize the weights and biases from the input layer to the hidden layer. Through the alternating iterative optimization of the two - layer parameters, the overall parameters that can minimize the operating goal are obtained.

[0032] 2.2 Construct a multi - layer parameter optimization mathematical model, including the solution of two - layer parameters from the input layer to the hidden layer and from the hidden layer to the output layer, specifically including:

[0033] i) For any source - load power prediction, there are N training samples at N time points in the training set Where: X i =[x i1 ,x i2 ,…,x in ∈R n is the input feature at each time point, and y i ∈R m is the true value corresponding to the training. The input neurons in the feature mapping process, that is, the number of predicted features is n, the number of hidden - layer neurons is L, the number of output neurons is 1, and the activation function g(x) is set as the sigmoid function, specifically: w i =[w i1 ,w i2 ,…,w in T is the input weight connecting the i - th hidden - layer neuron and the input - layer neuron, b i is the bias of the i - th hidden - layer neuron, β = [β1,β2,…,β L T is the weight matrix between the hidden layer and the output layer, and H is the matrix of the hidden layer after random feature mapping.

[0034] Since the neuron parameters (w i ,b i ​​)After randomly generating and giving training samples according to the probability of any continuous sampling distribution, the output matrix H of the hidden layer is actually known and remains unchanged, that is, the weight matrix β between the hidden layer and the output layer is to be solved.

[0035] ii) The day-ahead and intra-day two-stage operation cost objectives of the park integrated energy system are: Where: and represent the charging and discharging powers of the electricity storage device at time t, represents the electricity purchase quantity from the external power grid for day-ahead dispatch optimization, represents the power of the diesel generator at time t for intra-day dispatch optimization, represent the outputs of the gas boiler and gas turbine at time t obtained from intra-day stage optimization respectively.

[0036] iii) Add the weight matrix β to the optimization model and jointly optimize it with the other decision variables: z,β=argminF(z ID ,z DA (β);Y), where: z DA is the day-ahead decision variable determined according to the outputs of all prediction models and day-ahead constraint conditions during day-ahead optimization, and z ID is the intra-day decision variable determined according to z DA and the actual source-load power y during intra-day optimization.

[0037] iv) Find the corresponding β with the goal of minimizing the operation cost F of the park integrated energy system on the training set, and the predicted value on the test set is obtained by multiplying the matrix H after random feature mapping from the input layer to the hidden layer and the weight matrix β from the hidden layer to the output layer:

[0038] Through specific practical experiments, in the specific environment of a campus integrated energy system consisting of wind power, photovoltaic power, electrical load, and thermal load, a multi-layer prediction parameter optimization method was initiated using a genetic algorithm (population size 50, evolutionary generation number 40, two-point crossover probability 0.5, Gaussian mutation standard deviation 0.5 / mutation probability 0.2, and tournament selection size 3). A parameter iteration mechanism was coupled with a two-stage operation strategy (day-ahead and intraday optimization) using a sigmoid activation function hidden layer feature mapping. The experimental data obtained showed that, driven by 7 days of historical data from the training set, the total operating cost of the test set was reduced to US$11,363.5 by the 31st generation (a 0.15% reduction compared to the single-layer linear regression method), and the RMSE errors for wind, photovoltaic, and thermal load predictions were 1.999, 0.644, 1.244, and 0.063, respectively. By adjusting the objective function error weight parameter λ, the dynamic balance between operating cost and prediction accuracy was verified. As λ increased, the operating cost gradually increased from US$11,363.5, while the RMSE for electrical load decreased to 0.069. In addition, the forecast deviation guided by the operation strategy is manifested as a systematic low forecast value on the load side and a high forecast value on the new energy side, in order to optimize the consumption of new energy and reduce the total cost of the system.

[0039] Compared with the existing technology, this method solves the problem that traditional linear regression models are difficult to represent complex relationships such as spatiotemporal coupling nonlinearity and multi-energy flow interaction nonlinearity by introducing nonlinear feature mapping (Sigmoid activation function) and a two-layer parameter optimization framework driven by genetic algorithm (population size 50, evolutionary generations 40). The method reduces the RMSE error of wind, solar and thermal prediction by 48.2% (photovoltaic from 1.243 to 0.644), 44.2% (wind power from 1.064 to 1.999) and 63.2% (thermal load) respectively compared with the single-layer linear model. At the same time, by coupling a two-stage operation strategy (day-ahead-intraday optimization) with a multi-objective balancing mechanism (λ parameter control), the total operating cost of the test set was further reduced to US$11,363.5 (a 0.15% reduction compared to single-layer optimization), achieving dynamic adaptation of economy and accuracy (for example, the RMSE of the electric load can be reduced from 1.244 to 0.069 when λ is increased). Existing technologies, however, lack the ability to express nonlinear features and optimize global parameters, and are therefore unable to achieve both high-precision prediction and operating cost optimization.

[0040] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principles and purpose of the present invention. The scope of protection of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. All implementation schemes within its scope shall be subject to the constraints of the present invention.

Claims

1. A decision-making oriented multi-layer optimization prediction method for source and load integrating non-linear features, characterized in that After constructing and initializing the two-layer parameter optimization problem, the calculated hidden layer parameters are input into the two-stage optimization model. After obtaining the output layer parameters through the optimal operation cost objective optimization algorithm, the genetic algorithm is used to update and generate new hidden layer parameters, and the loop optimization is carried out until a two-layer parameter combination that achieves the optimal balance between the operation cost and the prediction error is obtained, that is, the non-linear mapping parameters of the hidden layer and the linear weight parameters of the output layer, so as to realize the optimized prediction; The two-layer parameter optimization problem mentioned above refers to the hierarchical optimization of the input layer to hidden layer parameters and the hidden layer to output layer parameters. The former is globally optimized by the genetic algorithm, and the latter is dynamically adjusted by the solver in combination with the operation strategy of the two-day-ahead and intraday optimization stages. The two are alternately iterated to achieve the global optimum; The two-stage optimization model mentioned above refers to a collaborative optimization framework that embeds the operation decisions of the two-day-ahead and intraday stages, and synchronously adjusts the prediction parameters and operation variables in the optimization of the hidden layer to output layer parameters to minimize the total operation cost.

2. The decision-making oriented multi-layer optimization prediction method for source and load integrating non-linear features according to claim 1, characterized in that, The optimal operation cost objective optimization algorithm mentioned above refers to the global search strategy based on the genetic algorithm. Taking the operation cost of the test set as the fitness function, the input layer parameters are optimized through the evolutionary mechanism, and the output layer parameters are optimized in combination with the two-stage solver, and finally the total operation cost is minimized.

3. The decision-making oriented source-load multi-layer optimization prediction method integrating non-linear features according to claim 1 or 2, characterized in that specifically Including: Step 1: Mapping the non-linear characteristics of the source and load of the campus integrated energy system, and using the non-linear characteristics to model and optimize the campus integrated energy system, specifically including: 1.1 Nonlinear transformation of prediction features: Wind power, photovoltaic power, electrical load, and thermal load in the integrated energy system of the park all have their own prediction features. By optimizing the prediction parameters, the linear model approximates the minimum objective of prediction accuracy or operating cost. Specifically: Y n = θ n X n , n ∈ {WT, PV, Eload, Hload}, where: Y n and θ n are vectors representing the source-load prediction results and prediction parameters on the representative dataset, with a dimension of N×1, where N is the number of samples, and X n is the matrix of prediction features, the input feature matrix, with a dimension of N*T*k n , T is the time scale, and k n is the number of features (e.g., wind power includes features such as wind speed and wind direction); the feature parameters Y n = θ' n ·H n = θ' n ·g(w n X n + b n ), n ∈ {WT, PV, Eload, Hload}, H n is the hidden layer feature matrix after feature transformation, θ' n is the parameter of the transformed feature, and g(·) is the Sigmoid or ReLU function; 1.2 Optimization of the input layer parameters based on the genetic algorithm; Step 2: Step-by-step optimization of the multi-layer prediction parameters from the input layer to the hidden layer and from the hidden layer to the output layer, specifically including: 2.1 In the source and load power prediction oriented by the operation cost, the parameters from the input layer to the hidden layer and from the hidden layer to the output layer need to be solved separately. Specifically, for the parameters from the input layer to the hidden layer, the genetic algorithm is used to select the optimal parameters. After each given set of parameters from the input layer to the hidden layer, the solver is used to solve the optimization problem of the hidden layer to output layer parameters with the embedded two-stage operation strategy, and the corresponding operation cost under the optimal parameters is obtained; then, with the goal of minimizing the operation cost, the genetic algorithm is used to optimize the weights and biases from the input layer to the hidden layer. Through the alternating iteration optimization of the two-layer parameters, the overall parameters that can minimize the operation goal are obtained; 2.2 Construct a multi-layer parameter optimization mathematical model, including the solution of the two-layer parameters from the input layer to the hidden layer and from the hidden layer to the output layer.

4. The decision-making oriented multi-layer optimization prediction method for source and load integrating non-linear features according to claim 3, characterized in that, The genetic algorithm mentioned above specifically includes: i) For the source-load power in the park, its input feature is X n , n ∈ {WT, PV, Eload, Hload}. To use a matrix for both the weights and biases corresponding to the features simultaneously, add a column of all 1s to the end of the input feature: X' n = [X n 1]. To optimize the source-load features simultaneously, let the input layer X' = [X' WT , X' PV , X' Eload , X' Hload , that is, create a matrix of size N*T*(k WT + k PV + k Eload + k Hload + 4) to store the weights and bias values of the features for source-load power prediction after mapping, where: k WT , k PV , k Eload , k Hload are the respective numbers of corresponding features of WT, PV, Eload, and Hload; ii) Transform the k all features corresponding to the bias and the k all +4 inputs at each moment into L features on the hidden layer, obtaining the matrix H = g(AX') of the hidden layer after the random feature mapping, where: g(·) is a non-linear activation function, A is a matrix of size (k all +4)*L, and the total number of features k all = k WT + k PV + k Eload + k Hload ; iii) Initialize the population: Set the population size to 50. For each type of source and load characteristic to be mapped, construct a vector with a length of (k + 1)*L, and assign the elements in it to real numbers between [-1, 1], and then perform selection operations, crossover operations, and mutation operations in turn; iv) For the individuals that have experienced step iii, calculate their population fitness. In general optimization problems, fitness is the objective function. However, for the double-layer parameter optimization problem, the selected individuals need to be transferred back to the matrix A from the input layer to the hidden layer. The corresponding hidden layer matrix is H. For the parameters from the output layer to the hidden layer, first, the weight matrix β is solved according to the generalized inverse matrix theory as the initial value of the convex optimization with the two-stage operation objective embedded. Then, a solver is called to optimize the operation problem containing β and the day-ahead and intra-day two-stage decision variables to obtain a better parameter β' and the corresponding operation cost C of the test set. v) Set the number of generations of evolution to 40. Use the operation cost C of the test set as the fitness function, and repeat the crossover, mutation, and selection operations with the goal of minimizing the fitness function. After multiple generations of evolution, the individuals in the population approach the optimal solution.

5. The decision-making oriented multi-layer optimization prediction method for source and load integrating non-linear features according to claim 4, characterized in that The selection operation mentioned above, namely tournament selection, sets the tournament size to 3. When performing the tournament operation, first, for each selection process, randomly select 3 individuals from the population. Then, compare the fitness of these 3 individuals. Finally, select the individual with the highest fitness among these 3 individuals and put it into the next-generation population. Repeat this process until the number of individuals in the new population reaches the population size of 50. The crossover operation mentioned above, namely the two-point crossover method, has a crossover probability of 0.

5. The mutation operation, i.e., the Gaussian mutation method that perturbs the individual genes based on the Gaussian distribution (normal distribution), randomly changes the genes of the individual. Assume that the individual X to be mutated in the population is X = [x1, x2, …, x n , where: x i (i = 1, 2, …, n) represents the gene. Then, for each gene x i , a random number Δx 2 that follows the Gaussian distribution N(0, σ i ) will be generated. This random number is the perturbation amount of the mutation. The result of the Gaussian mutation is:

6. The decision-making oriented multi-layer optimal prediction method for source and load integrating non-linear features according to claim 1, characterized in that Step 2.2 mentioned above specifically includes: i) Assume that for any source-load power prediction, there are N training samples at N moments in the training set Where: X i =[x i1 ,x i2 ,…,x in ∈R n is the input feature at each moment, y i ∈R m is the true value corresponding to the training. The input neurons in the feature mapping process, that is, the number of predicted features is n, the number of neurons in the hidden layer is L, the number of output neurons is 1, and the activation function g(x) is set to the sigmoid function. Specifically: Where: w i =[w i1 ,w i2 ,…,w in T is the input weight connecting the i-th hidden layer neuron and the input layer neuron, b i is the bias of the i-th hidden layer neuron, β = [β1, β2, …, β L T is the weight matrix between the hidden layer and the output layer, and H is the matrix of the hidden layer after random feature mapping;​​ ii) The day-ahead and intra-day two-stage operating cost targets of the park integrated energy system are as follows: Among them: and respectively represent the charging and discharging powers of the electricity storage equipment at time t, represents the electricity purchase quantity from the external power grid for day-ahead dispatch optimization, represents the power of the diesel generator at time t for intra-day dispatch optimization, respectively represent the outputs of the gas boiler and gas turbine at the optimized time in the intra-day stage; iii) Incorporate the weight matrix β into the optimization model and jointly optimize it with the remaining decision variables: z,β = argminF(z ID ,z DA (β); Y), where: z DA is the day-ahead decision variable determined according to the outputs of all prediction models and the day-ahead constraint conditions during day-ahead optimization, and z ID is the intra-day decision variable determined according to z DA and the actual source-load power y during intra-day optimization; iv) Taking the minimum operation cost F of the integrated energy system in the park on the training set as the goal, the corresponding β is obtained. The predicted value on the test set is obtained by multiplying the matrix H after random feature mapping from the input layer to the hidden layer and the weight matrix β from the hidden layer to the output layer:

7. A decision-making oriented source-load multi-layer optimization prediction system integrating non-linear features for implementing the method according to any one of claims 1-6, characterized in that, including: a feature mapping and non-linear transformation unit, a genetic algorithm parameter optimization unit, a two-stage operation decision unit, and a multi-objective balance regulation unit. Among them: the feature mapping and non-linear transformation unit performs non-linear feature mapping processing on the original feature information of the source-load power through an activation function to obtain a hidden layer feature matrix; the genetic algorithm parameter optimization unit performs global optimization of the weights and biases from the input layer to the hidden layer through the genetic algorithm according to the hidden layer feature matrix and the feedback information of the test set operation cost to obtain the optimal input layer parameters; the two-stage operation decision unit performs collaborative optimization of the parameters and operation variables from the hidden layer to the output layer through an embedded two-stage optimization model according to the hidden layer feature matrix and the day-ahead and intra-day energy demand information to obtain the prediction parameters and scheduling plan that minimize the total operation cost; the multi-objective balance regulation unit performs weight allocation processing of the operation cost and prediction error by dynamically adjusting the error weight parameter λ in the objective function according to the economic and accuracy requirement information set by the user to obtain an optimized parameter combination suitable for different scenarios and realize the flexible trade-off between system economy and prediction accuracy.

Images