Power prediction method fusing improved snow ablation optimization algorithm and CNN+KAN+LSTM + MSA hybrid model

Abstract

The invention discloses a hybrid model power prediction method fusing an improved snow ablation optimization algorithm and CNN + KAN + LSTM + MSA. The method comprises the following steps: constructing a snow ablation optimization algorithm according to a fusion good point set strategy and a periodic oscillation mutation strategy; and a CNN-KAN + BiLSTM-MSA hybrid model fused with the GVSAO algorithm and a hyper-parameter learning process are constructed. Through a good point set strategy, the initial solution distribution is more uniform, the coverage solution space is wider, and the global search capability is enhanced; the periodic oscillation mutation strategy prevents premature convergence of the solution, keeps diversity of the solution, and ensures exploration of the globally optimal solution; in the local development stage, candidate solutions are refined step by step, and the understanding quality and convergence efficiency are improved; the performance of a snow ablation optimization algorithm is greatly improved, and a global optimal solution can be better found; spatial feature extraction and time sequence learning are combined, so that the problem of time sequence prediction is effectively solved.

Description

Technical Field

[0001] The present invention relates to a power generation prediction method, and in particular to a power prediction method that integrates an improved snow melting optimization algorithm with a CNN+KAN+LSTM+MSA hybrid model. Background Art

[0002] In recent years, with the continuous development of wind power technology, the global wind power market and the scale of grid-connected wind farms have also been expanding. However, the uncertainty and instability of wind farm power caused by the randomness and volatility of wind energy have brought severe challenges to large-scale wind power grid connection. In order to reduce the impact of wind power grid connection and improve the stability of the power system, it is particularly important to effectively control wind turbines while meeting the power supply demand and carry out wind power prediction research. Existing prediction technologies face challenges in processing highly nonlinear and uncertain data, and also have difficulties in expressing explanatory insights, such as long training time and high model complexity. Especially on large-scale data sets, their computational efficiency is low and the overall performance of the algorithm is poor. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a power prediction method by integrating the improved snow melting optimization algorithm CNN+KAN+LSTM+MSA hybrid model.

[0004] Technical solution: The present invention comprises the following steps:

[0005] (1) Construct a snow melting optimization algorithm based on the fusion of the best point set strategy and the periodic oscillation mutation strategy;

[0006] (2) Construct the CNN-KAN+BiLSTM-MSA hybrid model integrating the GVSAO algorithm and the hyperparameter learning process.

[0007] Furthermore, the step (1) comprises:

[0008] (1.1) Combine the good point set strategy to generate the initial solution;

[0009] (1.2) Global exploration and preliminary search;

[0010] (1.3) Simulate the snow melting process and further improve the solution through local search;

[0011] (1.4) Introducing periodic oscillation mutation strategy;

[0012] (1.5) Convergence judgment and optimization termination.

[0013] Furthermore, in step (1.3), each candidate solution X best , and its local development is carried out through the following update formula:

[0014] X new =X best +α·(X best -X neighbor )

[0015] Among them, X neig hbor is the neighboring solution of the candidate solution (generated by random perturbation); α is a scaling factor that controls the step size of the local search and is between [0, 1].

[0016] Furthermore, the periodic oscillation mutation in step (1.4) is expressed as:

[0017] X mutate =X best +β·sin(w·t)+γ·rand(-1,1)

[0018] Among them, β is the oscillation hat degree, which controls the amplitude of the solution change; w is the oscillation frequency, which determines the periodicity of the mutation; t is the current number of iterations; γ is the mutation amplitude factor, which controls the intensity of the mutation; rand(-1, 1) is a random number between [-1, 1].

[0019] Furthermore, the convergence condition of step (1.5) is expressed as:

[0020] Δf=|f(X best )-f(X new )|<∈

[0021] Among them, ∈ is the set convergence threshold. When the change of the objective function is less than ∈, the algorithm is considered to have converged.

[0022] Furthermore, the step (2) comprises:

[0023] (2.1) Construct the model structure, including convolutional neural network, KAN network, bidirectional gated recurrent unit and multi-head self-attention mechanism;

[0024] (2.2) Combine the convolutional neural network, KAN network, bidirectional gated recurrent unit and multi-head self-attention mechanism components to build the entire model;

[0025] (2.3) The GVSAO algorithm is introduced to optimize the hyperparameters of CNN and BiLSTM networks, including initial solution generation, global exploration, local development, periodic oscillation mutation strategy, convergence judgment, data preparation and model training.

[0026] Furthermore, the bidirectional gated recurrent unit in step (2.1) includes an input layer and a bidirectional gated recurrent unit BiLSTM layer; the input layer is used to receive features extracted from KAN; the bidirectional gated recurrent unit BiLSTM layer is a bidirectional LSTM layer, which is used to capture the front-end dependency of the sequence.

[0027] Furthermore, the multi-head self-attention mechanism in step (2.1) includes an input layer and an MSA layer;

[0028] The input layer is used to receive the features output from BiLSTM;

[0029] The MSA layer is used for the multi-head self-attention mechanism to flexibly weight the input information at different time steps;

[0030]

[0031] Among them, Q, K, W are the input features after linear transformation, d k is the scaling factor.

[0032] Furthermore, the convergence judgment condition of step (2.3) is:

[0033] If|f(x new )-f(x current )|<∈ort≥t max , then stop.

[0034] The algorithm stops when the objective function converges or the maximum number of iterations is reached.

[0035] Furthermore, the model training in step (2.3) includes:

[0036] Initialize the model: Use the CNN-KAN-BiLSTM-MSA architecture;

[0037] Optimization model: Use GVSAO to optimize the model convolutional neural network, bidirectional gated recurrent unit BiLSTM hyperparameter parameters, and use Bayesian optimization KAN network.

[0038] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the present invention integrates the good point set strategy and the periodic oscillation mutation strategy into the snow melting optimization algorithm. Through the good point set strategy, the initial solution distribution is more uniform, covering a wider solution space, and enhancing the global search capability; the periodic oscillation mutation strategy prevents the solution from converging prematurely, maintains the diversity of solutions, and ensures the exploration of the global optimal solution; the candidate solutions are gradually refined in the local development stage to improve the quality of the solution and the convergence efficiency; the performance of the snow melting optimization algorithm is greatly improved, and the global optimal solution can be better found; the combination of spatial feature extraction and time series learning effectively solves the problem of time series prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a model structure diagram of the present invention;

[0040] Figure 2This is the fitness simulation result diagram of MSAO-CNN-KAN-BiLSTM-MSA;

[0041] Figure 3 It is the feature extraction map of MSAO;

[0042] Figure 4 is the loss value of the combined model;

[0043] Figure 5 This is the prediction result diagram of the MSAO-CNN-KAN-BiLSTM-MSA model. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0045] like Figure 1 As shown, the present invention provides a power prediction method integrating an improved snow melting optimization algorithm with a CNN+KAN+LSTM+MSA hybrid model, comprising the following steps:

[0046] (1) Combining the good point set strategy and periodic oscillation mutation strategy to construct a snow melting optimization algorithm

[0047] (1.1) Generation of initial solution (combined with good point set strategy)

[0048] A. Introduction of the good point set strategy

[0049] The good point set strategy generates an initial solution set through uniform sampling, ensuring the diversity and wide coverage of the initial solution. The good point set usually refers to a sampling method based on number theory, such as Latin hypercube sampling or uniform distribution sampling, which ensures that the distribution of solutions in the solution space is more uniform rather than randomly generated.

[0050] B. Formula description:

[0051] Assume that the solution space is D, its dimension is d, and the number of solutions is N. The goal of the good point set strategy is to generate a set of uniformly distributed initial solutions x0, each solution x i satisfy:

[0052] x i ={x i1 , x i2 , …, x id}, x ij ∈[L j , U j ]

[0053] Among them [L j , U j ] are the upper and lower bounds of the j-th dimension of the solution. Through the good point set strategy, the initial solution set of x0 is distributed as evenly as possible in the entire solution space.

[0054] (1.2) Global Exploration Phase

[0055] A. Global Exploration and Preliminary Search

[0056] The initial solution x0 is globally explored in the solution space, and each solution is evaluated according to the objective function f(X). The purpose of this stage is to explore the entire solution space and avoid falling into a local optimum at an early stage.

[0057] B. Formula description:

[0058] The objective function value f(X i ) is calculated as follows:

[0059] f(X i )=f(x i1 , x i2 , …, x id )

[0060] By evaluating the objective function, a part of the better solutions are selected as candidate solutions X best , these candidate solutions will enter the next stage of local development process.

[0061] (1.3) Local development stage (snow melting process)

[0062] A. Ablation strategy (local optimization)

[0063] In this phase, the algorithm simulates the melting process of snow and further improves the solution through local search. The ablation strategy makes small local adjustments to the candidate solutions and gradually narrows the search range.

[0064] B. Formula description:

[0065] For each candidate solution X best , and its local development is carried out through the following update formula:

[0066] X new =X best +α·(X best -X neig hbor )

[0067] Among them, X neig hbor is the neighboring solution of the candidate solution (generated by random perturbation); α is a scaling factor that controls the step size of the local search, usually between [0, 1]. This process is equivalent to performing a fine-grained search near the current solution to improve the quality of the solution.

[0068] (1.4) Introduction of periodic oscillation mutation strategy

[0069] A. Mutation Mechanism

[0070] In order to prevent the algorithm from falling into the local optimum, mutations are introduced periodically. The mutation strategy simulates the sudden jump of the solution, allowing the candidate solution to jump out of the local area and re-explore other areas.

[0071] B. Formula for periodic oscillation mutation:

[0072] After a certain period T, a mutation is applied to the solution:

[0073] X mutate =X best +β·sin(w·t)+γ·rand(-1,1)

[0074] Among them, β is the oscillation cap, which controls the change amplitude of the solution; w is the oscillation frequency, which determines the periodicity of the mutation; t is the current number of iterations; γ is the mutation amplitude factor, which controls the strength of the mutation; rand(-1, 1) is a random number between [-1, 1]. This formula makes the solution have both an oscillating change trend (simulated by the sine function) and a random jump effect (introducing mutations through random numbers) after period T.

[0075] (1.5) Convergence judgment and optimization termination

[0076] A. Convergence Conditions

[0077] The optimization process is terminated when the algorithm meets any of the following convergence conditions: the change in the objective function value is lower than a certain threshold; the maximum number of iterations N is reached max ; The diversity of solutions decreases to a certain extent.

[0078] B. Formula description:

[0079] The convergence condition can be expressed as:

[0080] Δf=|f(X best )-f(X new )|<∈

[0081] Where ∈ is the set convergence threshold. When the change of the objective function is less than ∈, the algorithm is considered to have converged. From the above improved algorithm steps, it can be seen that in the initialization process, the introduction of the optimal point set initialization strategy and the uniform distribution of the population can increase the diversity of the population and help to get rid of the attraction of the local minimum. The SAo algorithm introduces a dual population mechanism, which divides the population into two sub-populations, responsible for exploration and utilization, to achieve balance in the solution space; in the exploration stage, the algorithm uses Brownian motion to simulate the irregular movement of steam, so that the search agent has a highly dispersed feature, thereby exploring potential areas; in the utilization stage, SAo uses the degree-day method to simulate the snowmelt process, encouraging the search agent to find high-quality solutions near the current best solution; the SAO algorithm uses an overall position update equation, combined with exploration and utilization strategies, and population dynamic changes to optimize the position of each individual. In the SAO optimization process, a periodic oscillation mutation strategy is designed to prevent the solution from converging prematurely, maintain the diversity of the solution, and ensure the exploration of the global optimal solution. Therefore, this fusion strategy can theoretically greatly improve the performance of the snowmelt optimization algorithm, especially in complex multi-peak optimization problems, and can better find the global optimal solution.

[0082] (2) Constructing the CNN-KAN+BiLSTM-MSA hybrid model integrating the GVSAO algorithm and the hyperparameter learning process

[0083] The 24-step-ahead multivariate time series regression prediction algorithm based on the GVSAO-optimized CNN-KAN-BiLSTM-MSA model adopts and optimizes the following four main components: Convolutional Neural Network (CNN) is used to process multi-channel inputs of multivariate time series and capture the spatial relationship between input features; KAN network uses Bayesian to find the optimal KAN single-dimensional function form and quantity; BiLSTM is used to capture long-distance dependencies in the sequence and consider past and future information simultaneously through bidirectionality; Multi-head self-attention mechanism (MSA) enables the model to more flexibly weight input information at different time steps, handle long-term dependencies, and enhance the perception of global information. Combined with the snow melting optimization algorithm (GVSAO) based on the fusion of good point set strategy and periodic oscillation mutation strategy designed in this paper, the model parameters are optimized to improve the prediction performance. The following are the detailed steps:

[0084] (2.1) Model structure

[0085] A. Convolutional Neural Network (CNN)

[0086] Input layer: multivariate time series data, shape is

[0087] (batch_size, num_channels, sequence_length)

[0088] X∈R batc h_size×num-channels×sequence_length

[0089] Convolutional layer: Multiple convolutional layers extract spatial features. For the lth layer of convolution operation;

[0090] H (l) =f(W (l) *H (l-1) +b (l) )

[0091] Among them, H (0) =X,W (l) and b (l) are the convolution kernel and bias of the lth layer respectively, * represents the convolution operation, and f is the activation function such as ReLU.

[0092] Pooling layer: downsample the convolutional features. Let the pooling function be P,

[0093] H pool =P(H (L) )

[0094] Where L represents the last convolutional layer.

[0095] B.KAN Network

[0096] Input layer: receives features extracted from CNN.

[0097] Z=H pool

[0098] Bayesian Optimization:

[0099] Define the objective function f(x).

[0100] Construct a Gaussian process model:

[0101] The Gaussian process model is updated iteratively based on the new evaluation results.

[0102] Output layer: Process the input features using the optimal single-dimensional function combination and quantity after Bayesian optimization.

[0103] K=KAN(Z)

[0104] C. Bidirectional Gated Recurrent Unit (BiLSTM)

[0105] Input layer: receives features extracted from KAN.

[0106] G = KBiLSTM layer: bidirectional LSTM layer, capturing the dependencies between the sequences.

[0107]

[0108]

[0109]

[0110] D. Multi-head self-attention mechanism (MSA)

[0111] Input layer: receives the features output from BiLSTM.

[0112] M=H

[0113] MSA layer: Multi-head self-attention mechanism, flexibly weighting input information at different time steps.

[0114]

[0115] Among them, Q, K, W are the input features after linear transformation, d k is the scaling factor.

[0116] (2.2) Combination model

[0117] Combine the above components to build the entire model:

[0118] Model=MSA(BiGRU(KAN(CNN(X))))

[0119] (2.3) Introducing the GVSAO algorithm to optimize CNN and BiLSTM network hyperparameters

[0120] A. Initial solution generation (good point set strategy)

[0121] Generate uniformly distributed initial solutions through the good point set strategy.

[0122] X0={x1,x2,...,x n}

[0123] Among them, x i represents the initial solution, which is randomly generated using a uniform distribution.

[0124] B. Global Exploration

[0125] Perform a global search on the initial solution, evaluate the objective function, and select the better solution to enter the next stage.

[0126] f(x)=MAPE(y true ,y pred )

[0127] Here, MAPE stands for mean absolute percentage error.

[0128] C. Local development (ablation process)

[0129] Perform local search in the neighborhood of candidate solutions and refine the solution through progressive ablation.

[0130] X new =X current +α·random(-1,1)

[0131] Among them, α is the step size coefficient.

[0132] D. Periodic Oscillation Mutation Strategy

[0133] Periodically introduce oscillation mutations to break the local optimum and re-explore underdeveloped areas.

[0134] X mutated =X+β·sin(γt)

[0135] Where β is the amplitude, γ is the frequency, and t is the number of iterations.

[0136] E. Convergence Judgment

[0137] The algorithm stops when the objective function converges or the maximum number of iterations is reached.

[0138] If|f(x new )-f(x current )|<∈ort≥t max , then stop.

[0139] (2.4) Training and prediction

[0140] A. Data Preparation

[0141] Prepare training and test sets for multivariate time series data.

[0142] (X train ,y train ), (X test ,y test )

[0143] B. Model Training

[0144] Initialize the model: Use the CNN-KAN-BiLSTM-MSA architecture.

[0145] Optimize the model: Use GVSAO to optimize the model CNN and BiLSTM hyperparameters, and use Bayesian optimization KAN.

[0146] Train Model on(X train ,y train )

[0147] Evaluate Model on(X test ,ytest )

[0148] Through these detailed mathematical formulas and steps, the GVSAO optimization algorithm can effectively optimize the hyperparameters of the CNN-KAN-BiLSTM-MSA model and improve the performance of multivariate time series regression prediction.

[0149] like Figure 2 → Figure 5 The figure shows the simulation results after the improvement of the snow melting optimization algorithm of the present invention, where: Figure 2 This is the fitness simulation result diagram of MSAO-CNN-KAN-BiLSTM-MSA; Figure 3 It is the feature extraction map of MSAO; Figure 4 is the loss value of the combined model; Figure 5 This is the prediction result diagram of the MSAO-CNN-KAN-BiLSTM-MSA model.

Claims

1. A power prediction method integrating an improved snow melting optimization algorithm with a CNN+KAN+LSTM+MSA hybrid model, characterized in that: The steps include: (1) Construct a snow melting optimization algorithm based on the fusion of the best point set strategy and the periodic oscillation mutation strategy; (2) Construct the CNN-KAN+BiLSTM-MSA hybrid model integrating the GVSAO algorithm and the hyperparameter learning process.

2. According to claim 1, the power prediction method of the fusion improved snow melting optimization algorithm and the CNN+KAN+LSTM+MSA hybrid model is characterized in that: The step (1) comprises: (1.1) Combine the good point set strategy to generate the initial solution; (1.2) Global exploration and preliminary search; (1.3) Simulate the snow melting process and further improve the solution through local search; (1.4) Introducing periodic oscillation mutation strategy; (1.5) Convergence judgment and optimization termination.

3. The power prediction method based on the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 1 is characterized in that: In step (1.3), each candidate solution X best , and its local development is carried out through the following update formula: X new =X best +α·(X best -X neighbor ) Among them, X neighbor is the neighboring solution of the candidate solution (generated by random perturbation); α is a scaling factor that controls the step size of the local search and is between [0, 1].

4. The power prediction method of the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 1 is characterized in that: The periodic oscillation mutation in step (1.4) is expressed as: X mutate =X best +β·sin(w·t)+γ·rand(-1,1) Among them, β is the oscillation hat degree, which controls the amplitude of the solution change; w is the oscillation frequency, which determines the periodicity of the mutation; t is the current number of iterations; γ is the mutation amplitude factor, which controls the intensity of the mutation; rand(-1, 1) is a random number between [-1, 1].

5. The power prediction method of the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 1 is characterized in that: The convergence condition of step (1.5) is expressed as: Δf=|f(X best )-f(X new )|<∈ in, ∈ It is the set convergence threshold. When the change of the objective function is less than ∈, the algorithm is considered to have converged.

6. The power prediction method of the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 1 is characterized in that: The step (2) comprises: (2.1) Construct the model structure, including convolutional neural network, KAN network, bidirectional gated recurrent unit and multi-head self-attention mechanism; (2.2) Combine the convolutional neural network, KAN network, bidirectional gated recurrent unit and multi-head self-attention mechanism components to build the entire model; (2.3) The GVSAO algorithm is introduced to optimize the hyperparameters of CNN and BiLSTM networks, including initial solution generation, global exploration, local development, periodic oscillation mutation strategy, convergence judgment, data preparation and model training.

7. The power prediction method of the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 6 is characterized in that: The bidirectional gated recurrent unit in step (2.1) includes an input layer and a bidirectional gated recurrent unit BiLSTM layer; the input layer is used to receive features extracted from KAN; the bidirectional gated recurrent unit BiLSTM layer is a bidirectional LSTM layer, which is used to capture the front-end dependency of the sequence.

8. The power prediction method of the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 6 is characterized in that: The multi-head self-attention mechanism in step (2.1) includes an input layer and an MSA layer; The input layer is used to receive the features output from BiLSTM; The MSA layer is used for the multi-head self-attention mechanism to flexibly weight the input information at different time steps; Among them, QK, W are the input features after linear transformation, d k is the scaling factor.

9. The power prediction method of the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 6 is characterized in that: The convergence judgment condition of step (2.3) is: If|f(x new )-f(x current )|<∈ort≥t max ,thenstop. The algorithm stops when the objective function converges or the maximum number of iterations is reached.

10. The power prediction method of the fusion improved snow melting optimization algorithm and CNN+KAN+LSTM+MSA hybrid model according to claim 6 is characterized in that: The model training of step (2.3) includes: Initialize the model: Use the CNN-KAN-BiLSTM-MSA architecture; Optimization model: Use GVSAO to optimize the model convolutional neural network, bidirectional gated recurrent unit BiLSTM hyperparameter parameters, and use Bayesian optimization KAN network.

Images