Short-term electricity price prediction method and device based on crown porcupine optimization and deep learning

Abstract

The invention discloses a crown porcupine optimization and deep learning-based short-term electricity price prediction method and device, and belongs to the technical field of electricity price prediction, and the method comprises the steps: S1, carrying out the optimization of a VMD model based on crown porcupine optimization, obtaining an optimized VMD model, and decomposing first historical electricity price data into modal components based on the optimized VMD model; s2, introducing an index gating mechanism and a memory hybrid mechanism into LSTM to obtain an sLSTM model, and performing feature extraction on the modal component based on the sLSTM model to obtain a first time sequence feature vector representing a long-term dependency relationship; and S3, based on the first time sequence feature vector and second historical electricity price data, using a Transform model to obtain a second time sequence feature vector representing a short-term dependency relationship, and based on the first time sequence feature vector and the second time sequence feature vector, obtaining a short-term electricity price prediction result. The technical problems that nonlinear data are difficult to process and the long-term dependency relationship existing in the data is difficult to capture are solved, and the accuracy of electricity price prediction is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electricity price forecasting, and specifically to a short-term electricity price forecasting method and device based on crown porcupine optimization and deep learning. Background Art

[0002] With the rapid development and increasing complexity of the power system, electricity price forecasting plays a crucial role in optimizing power resource allocation, alleviating grid load pressure, improving power supply efficiency, and system operation reliability. On the load side, electricity price forecasting can guide users to actively adjust their electricity consumption behavior during peak load periods or when power supply is tight, promoting the efficient consumption of new energy and the widespread application of renewable energy. Therefore, in-depth research and development of efficient and accurate electricity price forecasting models have important theoretical value and practical significance for promoting the construction of a new power system and ensuring the stable operation of the power market. Existing technologies such as a method and system for forecasting the day-ahead electricity price in the power market based on CNN-LSTM-Attention with the patent number CN117522443A include dividing the historical daily electricity price information data after standardization processing into a training set and a test set; inputting the training set into the CNN-LSTM model for parameter optimization learning to obtain an optimized CNN-LSTM model; automatically calculating the weighted average of the output data of the optimized CNN-LSTM model through an attention mechanism to obtain the CNN-LSTM-Attention model; and forecasting the day-ahead electricity price on the prediction day through the CNN-LSTM-Attention model. The above solution ensures the accuracy of the electricity price forecasting result to a certain extent. However, CNN relies on linear convolution operations and is difficult to process non-linear data, and the LSTM used is also restricted by output range and other factors and is difficult to capture the long-term dependence relationships existing in the data. Summary of the Invention

[0003] Aiming at the technical problems that the existing technology is difficult to process non-linear data and difficult to capture the long-term dependence relationships existing in the data, the present invention provides a short-term electricity price forecasting method and device based on crown porcupine optimization and deep learning. The optimized VMD model after crown porcupine optimization decomposes the electricity price data into modal components. The characteristics of VMD for processing non-linear data are used to process non-linear data. At the same time, crown porcupine optimization is also carried out to avoid the problem of insufficient decomposition. By introducing an exponential gating mechanism and a memory mixing mechanism into LSTM to obtain the sLSTM model, the limitations of LSTM in terms of output range and information sharing are broken through. Furthermore, the sLSTM model is used to extract features from the modal components to obtain the first time series feature vector, realizing the capture of the long-term dependence relationships existing in the data, solving the technical problems of being difficult to process non-linear data and difficult to capture the long-term dependence relationships existing in the data, and significantly improving the accuracy of electricity price forecasting.

[0004] To solve the above technical problems, the present invention provides a short-term electricity price prediction method based on crested porcupine optimization and deep learning, including the following steps: S1: Optimize the VMD model based on crested porcupine optimization to obtain an optimized VMD model, and decompose the first historical electricity price data into modal components based on the optimized VMD model; S2: Introduce an exponential gating mechanism and a memory mixing mechanism into LSTM to obtain an sLSTM model, and extract features from the modal components based on the sLSTM model to obtain a first time series feature vector representing long-term dependence relationships; S3: Based on the first time series feature vector and the second historical electricity price data, use the Transformer model to obtain a second time series feature vector representing short-term dependence relationships, and obtain a short-term electricity price prediction result based on the first time series feature vector and the second time series feature vector.

[0005] After adopting the above technical solutions, the present invention has the following advantages: The processing of non-linear data is realized through the characteristics of VMD for processing non-linear data. At the same time, the VMD is optimized by the crested porcupine to avoid the problem of insufficient data decomposition. By introducing the exponential gating mechanism and the memory mixing mechanism into LSTM to obtain the sLSTM model, the limitations of the output range and information sharing suffered by LSTM are broken through. Furthermore, the sLSTM model is used to extract features from the modal components to obtain the first time series feature vector, realizing the capture of long-term dependence relationships existing in the data, solving the technical problems of difficult processing of non-linear data and difficult capture of long-term dependence relationships existing in the data, and significantly improving the accuracy of electricity price prediction; By inputting the first time series feature vector representing long-term dependence relationships and the second historical electricity price data into the Transformer model suitable for capturing short-term dependence, the second time series feature vector representing short-term dependence relationships can be fully and accurately obtained. Furthermore, electricity price prediction is carried out through the fully captured long-term and short-term dependence relationships, further improving the accuracy of electricity price prediction.

[0006] Preferably, in S1, the optimizing the VMD model based on crested porcupine optimization to obtain an optimized VMD model includes: randomly generating the positions of each individual in the initial population, where the position information of each individual in the initial population represents a set of parameters to be screened for optimizing the VMD model; Calculate the fitness of each individual in the initial population and store the optimal solution of the population. During the exploration stage of the crested porcupine optimization, guide each individual in the initial population to move towards the position corresponding to the optimal solution of the population, update the range of the initial population to obtain the range of the first population, and during the exploitation stage of the crested porcupine optimization, guide each individual in the range of the first population to move towards the position corresponding to the optimal solution of the population to update the range of the first population; Obtain the final parameters according to the range of the first population and the optimal solution of the population, and use the final parameters to replace the initial parameters in the VMD model to obtain the optimized VMD model.

[0007] Preferably, in S1, decomposing the first historical electricity price data into modal components based on the optimized VMD model includes: S1a: Introduce the convergence parameter into the optimized VMD model to obtain the augmented Lagrangian expression, iteratively update the initial modal components based on the augmented Lagrangian expression, and determine whether the initial modal components meet the preset conditions. If they meet, the initial modal components are the modal components; otherwise, update the convergence parameter based on Wiener filtering and execute S1a.

[0008] Preferably, in S2, introducing the exponential gating mechanism and the memory mixing mechanism into the LSTM to obtain the sLSTM model includes: Introduce the exponential gating mechanism into the activation value calculation mechanism of the LSTM, and introduce the memory mixing mechanism into the cell state update mechanism of the LSTM to obtain the sLSTM model.

[0009] Preferably, in S2, extracting features from the modal components based on the sLSTM model to obtain the first time series feature vector representing long-term dependence relationships includes: Use the output of the exponential activation function in the sLSTM model to eliminate the old information of the cell units in the sLSTM model, and at the same time control the new information stored in the cell units through the output of the exponential activation function, thereby updating the sLSTM model; Perform a logarithmic operation on the output of the updated sLSTM model, obtain the rated state and the normalized state according to the logarithmic operation result, introduce the rated state and the normalized state into the updated sLSTM model, and extract features from the modal components through the updated sLSTM model to obtain the first time series feature vector.

[0010] Preferably, extracting features from the modal components based on the sLSTM model to obtain the first time series feature vector representing long-term dependence relationships further includes: Obtain the initial output of the exponential activation function based on the characteristics of the modal components, based on the initial output, use the third historical electricity price data to obtain the process prediction result, and update the initial output based on the process prediction result and the actual prediction result to obtain the output of the exponential activation function.

[0011] Preferably, the output of the exponential activation function in the sLSTM model for eliminating the old information of the cell units in the sLSTM model includes: f = exp(Wx t + Rh t-1 + b); In the formula, f is the output of the forget gate in the sLSTM model; x t is the input value input to the forget gate at time t; W is the weight vector between x t and f; h t-1 is the hidden state at time t - 1; R is the recurrent weight between h t-1 and f, and b is the bias term.

[0012] Preferably, the S3 includes: S31: Mapping the first time series feature vector and the feature sequence corresponding to the second historical electricity price data to a high-dimensional space through a Transformer model to obtain a high-dimensional sequence, and adding a position encoding representing the time order to the high-dimensional sequence; S32: Obtaining attention weights based on the position encoding, and obtaining a second time series feature vector based on the attention weights.

[0013] Beneficial effects of this solution: The processing of non-linear data is realized through the characteristics of VMD for processing non-linear data. At the same time, the VMD is optimized through the crown porcupine, avoiding the problem of insufficient data decomposition. The sLSTM model is obtained by introducing an exponential gating mechanism and a memory mixing mechanism into the LSTM, breaking through the limitations of the output range and information sharing of the LSTM. Furthermore, the first time series feature vector is obtained by extracting the features of the modal components through the sLSTM model, realizing the capture of the long-term dependence relationship existing in the data, solving the technical problems of difficult processing of non-linear data and difficult capture of the long-term dependence relationship existing in the data, and significantly improving the accuracy of electricity price prediction; By inputting the first time series feature vector representing the long-term dependence relationship and the second historical electricity price data into the Transformer model suitable for capturing short-term dependence, the second time series feature vector representing the short-term dependence relationship can be fully and accurately obtained. Furthermore, electricity price prediction is carried out through the fully captured long-term dependence relationship and short-term dependence relationship, further improving the accuracy of electricity price prediction; By introducing a convergence parameter into the optimized VMD model and iteratively updating the initial modal components, and further, when it is determined that the initial modal components do not meet the conditions, the convergence parameter is updated based on Wiener filtering, further avoiding the problems of insufficient decomposition and overfitting of the first historical electricity price data on the basis of gradually approaching the optimal solution in the optimized VMD model; By using the output of the exponential activation function in the sLSTM model to eliminate the old information of the cell units in the sLSTM model and control the new information stored in the cell units, the sLSTM model is updated, realizing the elimination of information irrelevant to electricity price prediction and the retention of relevant information, thereby reducing the complexity of the sLSTM model and improving the electricity price prediction efficiency; By introducing the rated state and the normalized state into the updated sLSTM model, the problem of numerical overflow caused by the introduction of the exponential gating mechanism is avoided, thereby improving the robustness of the updated sLSTM model.

[0014] The present invention also provides a short-term electricity price prediction device based on the crowned porcupine optimization and deep learning, including a modal component acquisition module, a dependency acquisition module, and a short-term electricity price prediction result acquisition module; The modal component acquisition module is used to optimize the VMD model based on the crowned porcupine optimization to obtain an optimized VMD model, and decompose the first historical electricity price data into modal components based on the optimized VMD model; The dependency acquisition module is used to extract features from the modal components based on the sLSTM model to obtain a first time series feature vector representing long-term dependencies, and is also used to obtain a second time series feature vector representing short-term dependencies based on the first time series feature vector and the second historical electricity price data using the Transformer model; The short-term electricity price prediction result acquisition module is used to obtain a short-term electricity price prediction result based on the first time series feature vector and the second time series feature vector.

[0015] Preferably, it further includes an sLSTM model acquisition module, which is used to introduce an exponential gating mechanism and a memory mixing mechanism into the LSTM to obtain an sLSTM model.

[0016] Beneficial effects of this solution: By using the characteristic of VMD to process non-linear data, the processing of non-linear data is realized. At the same time, the crowned porcupine is used to optimize VMD to avoid the problem of insufficient data decomposition. By introducing an exponential gating mechanism and a memory mixing mechanism into the LSTM to obtain an sLSTM model, the limitations of the output range and information sharing of the LSTM are broken through. Then, the sLSTM model is used to extract features from the modal components to obtain a first time series feature vector, realizing the capture of long-term dependencies existing in the data, solving the technical problems of difficult processing of non-linear data and difficult capture of long-term dependencies existing in the data, and significantly improving the accuracy of electricity price prediction; By inputting the first time series feature vector representing long-term dependencies and the second historical electricity price data into a Transformer model suitable for capturing short-term dependencies, the second time series feature vector representing short-term dependencies can be fully and accurately obtained. Furthermore, by fully capturing the long-term and short-term dependencies, electricity price forecasting is carried out, further improving the accuracy of electricity price forecasting. Description of the Drawings

[0017] By reading the detailed description of the non-limiting embodiments with reference to the following drawings, other features, objectives, and advantages of the present invention will become more apparent. The drawings are only for the purpose of showing the preferred embodiments and are not considered as limiting the present invention. Moreover, throughout the drawings, the same reference numerals are used to represent the same components.

[0018] Figure 1 It is a flowchart of the short-term electricity price forecasting method based on the crowned porcupine optimization and deep learning of the present invention; Figure 2 It is a flowchart of optimizing VMD in the short-term electricity price forecasting method based on the crowned porcupine optimization and deep learning of the present invention; Figure 3 It is a spectrogram obtained by performing FFT transformation on the modal components in the short-term electricity price forecasting method based on the crowned porcupine optimization and deep learning of the present invention; Figure 4 It is a marginal spectrogram obtained by performing FFT transformation on the modal components in the short-term electricity price forecasting method based on the crowned porcupine optimization and deep learning of the present invention; Figure 5 It is a schematic structural diagram of the sLSTM model in the short-term electricity price forecasting method based on the crowned porcupine optimization and deep learning of the present invention; Figure 6 It is an overall framework diagram of the forecasting model in the short-term electricity price forecasting method based on the crowned porcupine optimization and deep learning of the present invention; Figure 7 It is a performance comparison diagram of different forecasting models in the short-term electricity price forecasting method based on the crowned porcupine optimization and deep learning of the present invention. Detailed Embodiments

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only the best embodiments of the present invention, only for explaining the present invention, and do not limit the protection scope of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0020] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict operations (or steps) as sequential processes, many of the operations (or steps) can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operations are completed, but it can also have additional steps not included in the drawings; the process can correspond to a method, function, procedure, subroutine, subprogram, and so on.

[0021] Embodiment 1: As Figure 1 shown, a short-term electricity price prediction method based on crested porcupine optimization and deep learning includes the following steps: S1: Optimize the VMD model based on crested porcupine optimization to obtain an optimized VMD model, and decompose the first historical electricity price data into modal components based on the optimized VMD model.

[0022] In S1, the optimizing the VMD model based on crested porcupine optimization to obtain an optimized VMD model includes: Randomly generate the positions of each individual in the initial population, where the position information of each individual in the initial population represents a set of parameters to be screened for optimizing the VMD model; Calculate the fitness of each individual in the initial population and store the population optimal solution. In the exploration stage of crested porcupine optimization, guide each individual in the initial population to move towards the position corresponding to the population optimal solution to update the initial population range to obtain the first population range. In the exploitation stage of crested porcupine optimization, guide each individual in the first population range to move towards the position corresponding to the population optimal solution to update the first population range; Obtain the final parameters according to the first population range and the population optimal solution, and use the final parameters to replace the initial parameters in the VMD model to obtain the optimized VMD model.

[0023] In S1, the decomposing the first historical electricity price data into modal components based on the optimized VMD model includes: S1a: Introduce a convergence parameter into the optimized VMD model to obtain an augmented Lagrangian expression, iteratively update the initial modal components based on the augmented Lagrangian expression, and determine whether the initial modal components meet the preset conditions. If so, the initial modal components are the modal components; otherwise, update the convergence parameter based on Wiener filtering and execute S1a.

[0024] The Crested Porcupine Optimization Algorithm, i.e., the CPO algorithm, is a new meta-heuristic algorithm that simulates the defensive behavior of crested porcupines when facing predators. By adjusting the individuals in the population, it speeds up the convergence rate while maintaining population diversity and avoiding falling into local minima. In the Crested Porcupine Optimization Algorithm, the population size is first maximized and then gradually decreased during the optimization process to simulate the aggregation and dispersion behaviors of crested porcupines when facing predators, and random values are generated using a normal distribution to simulate the selection behavior of predators. The Crested Porcupine Optimization Algorithm is suitable for solving complex parameter optimization problems in models and can find the global optimal solution in a relatively short time. The VMD model, i.e., the Variational Mode Decomposition model, is an adaptive signal decomposition method. It can decompose complex signals into a finite number of Intrinsic Mode Functions (IMFs), and each IMF contains a significant feature or pattern of the original signal. The core idea is to represent the original signal as a superposition of a series of band-pass signals with adaptive bandwidths, and these band-pass signals are represented by the envelopes of their corresponding analytic signals. The decomposition process is completed by solving a variational problem, and the goal is to optimize the central frequencies and bandwidths of different modal signals to maintain the inherent frequency band separation of the signal.

[0025] In this embodiment, before performing VMD (Variational Mode Decomposition), four parameters need to be preset: the number of IMF (Intrinsic Mode Function) components k, the penalty factor α, the noise tolerance τ, and the convergence accuracy ε. Among them, the influence of τ and ε on the decomposition result is relatively small, and they can usually be set to default values. However, if the values of k and α are not appropriate, it may lead to excessive or insufficient decomposition results. Therefore, in order to eliminate the potential influence of artificially setting VMD (Variational Mode Decomposition) parameters on the electricity price prediction result, the Crested Porcupine Optimization is adopted, and the sample entropy is used as the evaluation criterion to search for the global optimal combination of the k and α parameters in the Crested Porcupine Optimization. As Figure 2 shown, the specific optimization steps are as follows: Initialize the optimization algorithm parameters including the population size N, the maximum number of iterations T max , the number of independent variables m, the upper limit of the independent variable the lower limit of the independent variable the objective function f, the number of loops T, the convergence rate a, the trade-off percentage T between the third defense mechanism and the fourth defense mechanism f ; The position information of each individual, i.e., each porcupine, represents a set of parameters to be screened for optimizing the VMD (Variational Mode Decomposition) model. The position of each porcupine in the population is randomly generated using the following formula: In the formula, represents the position information of the i-th porcupine; is the lower limit of the independent variable; is the upper limit of the independent variable; is a random number in [0, 1]; After initializing the population, calculate the fitness of each individual in the population And store the optimal solution of the population, i.e., the global optimal solution and Then start the search in the exploration space. The search mechanism includes two stages: exploration and exploitation. The exploration stage reflects the exploration behavior of the porcupine. In this stage, the population has two defense strategies: vision and sound. For the defense strategy of vision, the normal distribution is used to generate random values to mathematically simulate these options. If these random values are between -1 and 1, it is encouraged to approach the CP. Otherwise, the predator will move away from the CP. The specific formula is as follows: In the formula, represents the position information of the i-th porcupine at the (t + 1)-th iteration; is the vector generated between the current porcupine and a randomly selected porcupine from the population, representing the position of the predator at iteration t; τ1 represents a random number based on the normal distribution; τ2 represents a random number in the interval [0, 1]; t is the current iteration number; r is a random number between [1, N]. For the defense strategy of sound, as shown in the following formula: In the formula, r1 and r2 represent two random numbers between [1, N]; τ3 represents a random number in the interval [0, 1]; represents the position of the predator; represents simulating three effects of sound on the distance between the predator and the porcupine. In the exploitation stage, it reflects the exploitation behavior of the porcupine. In this stage, the population has two defense strategies: smell and physical attack. For the defense strategy of smell, as shown in the following formula: In the formula, r3 represents a random number between [1, N]; represents the position information of the i-th porcupine at the t-th iteration; δ, γ t , respectively represent the parameter for controlling the direction, the defense factor, and the smell diffusion factor. The specific calculations are as shown in the following formula: In the formula, represents the objective function value of the i-th porcupine at the t-th iteration, and ε is a small value to avoid division by zero; represents a numerical vector randomly generated between [0, 1]; is used to simulate three situations that occur in this strategy: When The porcupine will stop spreading the smell because the predator will stop moving due to fear of the porcupine. Therefore, the distance between the predator and the porcupine remains unchanged; When Due to the predator being nearby, the porcupine will significantly emit the smell; When It is a combination of 0 and 1. The predator will keep a safe distance from the porcupine. Therefore, there is no need to widely release its odor. For the defensive behavior against physical attacks, during the attack, the two objects strongly merge, representing a one-dimensional inelastic collision, as shown in the following formula: In the formula, τ4 and τ5 represent random values between the interval [0, 1]; α is the convergence speed factor; represents the average force of the porcupine affecting the i-th predator, and its calculation is shown in the following formula: In the formula, m i represents the mass of the i-th individual at iteration t; represents the initial velocity of the i-th porcupine at iteration t; represents the final velocity of the i-th porcupine at the next iteration t + 1. The final parameters, that is, the optimal solution, are obtained through the above steps. The initial parameters in the VMD (Variational Mode Decomposition) model are replaced with the optimal solution to obtain the optimized VMD model.

[0026] In this embodiment, the convergence parameters are the quadratic penalty term α and the Lagrange multiplier λ, and the augmented Lagrangian expression is: The alternating direction multiplier algorithm is used to solve the saddle point of the augmented Lagrangian expression. By iteratively updating the initial mode components {u k} and the set of central frequencies {w k} corresponding to the initial mode components, and gradually advancing according to a specific formula, the optimal solution of the constrained variational model is finally obtained: In the formula, w represents the frequency; respectively represent the Fourier transform of; represents the residual amount obtained after Wiener filtering; represents the central frequency of the modal quantity power spectrum; at the same time, λ is updated through the following formula, where τ represents the noise tolerance parameter. According to the calculation result of the following formula, it is evaluated whether the preset condition, that is, the preset accuracy standard ε, is satisfied. The preset accuracy standard is flexibly set according to user requirements, so as to determine whether to terminate the iterative process: Finally, all decomposed modal components and their corresponding central frequencies can be obtained. These modal components together constitute the feature set of the signal, providing the necessary data basis for subsequent feature learning and state recognition work. At the same time, the characteristics of VMD for processing nonlinear data are used to process nonlinear data, and the porcupine optimization is also used to optimize VMD (Variational Mode Decomposition) to avoid the problem of insufficient data decomposition.

[0027] In this embodiment, the optimization range of the modal component k is [2, 12], and the optimization range of the penalty factor a is [500, 2500]. Other parameters that need to be set in the VMD (Variational Mode Decomposition) algorithm have less influence on the decomposition effect and are set to empirical values, that is, tau = 0, init = 1, DC = 0, ε = 10-7. When the porcupine optimization algorithm iterates to the 8th time, the fitness reaches the best value of 451.108, and the optimal parameters [k, a] are [8, 685]. After putting the electricity price data into the VMD optimized by CPO for decomposition, 8 IMF components are obtained. Each IMF is subjected to FFT transformation and the spectrogram and marginal spectrogram are drawn, as Figure 3 and Figure 4 shown. As Figure 3 can be seen, each IMF component has a significant amplitude in different frequency ranges and has good independence, which indicates that VMD successfully decomposes the signal into components in different frequency bands. High-frequency components (such as IMF1 and IMF2) represent the rapid fluctuations in the electricity price data, while low-frequency components (such as IMF7 and IMF8) represent long-term trends or periodic changes. As Figure 4 can be seen, it shows the amplitudes of each IMF component at different frequencies. It can be seen that the distribution of each component in frequency is relatively clear and there is no obvious frequency aliasing. From the perspective of amplitude, the amplitude distribution of each IMF component also shows a certain regularity. Some components have higher amplitudes in specific frequency ranges, indicating that these frequency components occupy a larger proportion of the energy in the original signal.

[0028] In this embodiment, to verify the superiority of the porcupine optimization in optimizing the parameters of the variational mode decomposition (VMD), the present invention conducts four groups of experiments based on different combinations of k (number of modes) and a (penalty factor). By comparing the model performance under different parameter combinations, the effectiveness of the optimization strategy of the porcupine optimization algorithm can be clearly demonstrated. The experimental results are shown in Table 1: Table 1. Comparison table of evaluation indexes for different decomposition parameters When k = 8 and a = 685, all evaluation indexes of the model show the best performance. At this time, MAE is 1.48, RMSE is 2.04, MAPE is as low as 0.06%, R 2Up to 0.97. Under this combination, the model not only has the smallest error but also the highest prediction accuracy, indicating that the CPO algorithm can accurately find the optimal parameters of VMD and significantly improve the accuracy of electricity price prediction. When keeping k = 8 unchanged and increasing a to 2000, all evaluation indicators of the model show a significant decline. The MAE and RMSE rise to 3.69 and 5.32 respectively, the MAPE increases to 0.15%, and R 2 drops to 0.83. This result shows that an overly large penalty factor will lead to overfitting of the model, reduce its adaptability to the volatility of electricity price data, and thus affect the prediction accuracy. Further, when keeping a = 685 unchanged and increasing k to 9, although the performance of the model is still better than the combination with a = 2000, compared with the optimal combination with k = 8, its MAE, RMSE, and MAPE all increase, and R 2 also drops. This indicates that when the penalty factor is fixed, increasing the number of modes does not always improve the model performance, but may instead introduce unnecessary complexity and affect the accuracy of the prediction results. Finally, when k = 9 and a = 2000, the performance of the model reaches the worst level among the four groups of experiments. The MAE and RMSE are 3.7 and 4.55 respectively, the MAPE is 0.17%, and R 2 is 0.87. This result further confirms the necessity of the CPO algorithm in optimizing the VMD parameters, that is, through reasonable parameter selection, the prediction performance of the model can be significantly improved.

[0029] S2: Introduce the exponential gating mechanism and the memory mixing mechanism into LSTM to obtain the sLSTM model, and extract features from the modal components based on the sLSTM model to obtain the first time series feature vector representing long-term dependencies.

[0030] In S2, the introducing the exponential gating mechanism and the memory mixing mechanism into LSTM to obtain the sLSTM model includes: introducing the exponential gating mechanism into the activation value calculation mechanism of LSTM, and introducing the memory mixing mechanism into the cell state update mechanism of LSTM to obtain the sLSTM model.

[0031] In S2, the extracting features from the modal components based on the sLSTM model to obtain the first time series feature vector representing long-term dependencies includes: Using the output of the exponential activation function in the sLSTM model to eliminate the old information of the cell units in the sLSTM model, and at the same time controlling the new information stored in the cell units through the output of the exponential activation function, thereby updating the sLSTM model; Take the logarithm of the output of the updated sLSTM model, obtain the rated state and the normalized state according to the result of the logarithmic operation, introduce the rated state and the normalized state into the updated sLSTM model, and extract the features of the modal components through the updated sLSTM model to obtain the first time series feature vector.

[0032] The method for extracting the first time series feature vector representing the long-term dependence relationship by extracting the features of the modal components based on the sLSTM model further includes: Obtain the initial output of the exponential activation function based on the characteristics of the modal components, obtain the process prediction result using the third historical electricity price data based on the initial output, and update the initial output based on the process prediction result and the actual prediction result to obtain the output of the exponential activation function.

[0033] The method for using the output of the exponential activation function in the sLSTM model to eliminate the old information of the cell units in the sLSTM model includes: f = exp(Wx t + Rh t-1 + b); In the formula, f is the output of the forget gate in the sLSTM model; x t is the input value input to the forget gate at time t; W is the weight vector between x t and f; h t-1 is the hidden state at time t-1; R is the recurrent weight between h t-1 and f, and b is the bias term.

[0034] LSTM, that is, the long short-term memory network, is a unique recurrent neural network. By integrating fine gating structures such as the input gate, forget gate, and output gate, it successfully overcomes the problems of gradient disappearance or gradient explosion that easily occur in traditional recurrent neural networks when processing long sequence data, enabling it to efficiently capture the long-term dependence characteristics existing in the sequence data.

[0035] In this embodiment, introducing the exponential gating mechanism and the memory mixing mechanism into the LSTM (Long Short-Term Memory Network) to obtain the sLSTM model is specifically as follows: In the input gate and the forget gate, the sLSTM introduces the exponential gating mechanism, and calculates the activation value of the gate through the exponential function, so that the output of the gate is no longer limited to between 0 and 1, but can be dynamically adjusted as needed. This mechanism allows a wider range of information to pass through, making the transmission of information more flexible and better capturing and utilizing the pattern of information flow; in the original LSTM (Long Short-Term Memory Network), the update of the memory cell state is achieved through a simple linear combination, which may cause information to be gradually lost or blurred during the transmission process. The memory mixing mechanism promotes the complex interaction and information transmission between memory units within the same head at different time nodes, enabling it to more effectively capture the long-term dependence characteristics in time series data, thereby improving the performance and prediction accuracy of the model.

[0036] In this embodiment, controlling the new information stored in the cell unit through the output of the exponential activation function includes: where i t is the output value of the forget gate; is the weight vector between x t and the input gate i t ; R i is the recurrent weight between h t-1 and the input gate i t ; b i is the bias term. Since using the exponential activation function may result in large values, causing overflow, an additional state m t is adopted in the sLSTM to stabilize the gating, and in the forward propagation, f t is replaced by f′ t , and i t is replaced by i′ t . The specific calculation method is as follows: m t = max(log(f t ) + m t-1 , log(i t )) f′ t = exp(log(f t ) + m t-1 - m t ) Regarding the calculation methods of the temporary memory cell z t and the output gate o t , they are still the same as those of the original LSTM (Long Short-Term Memory Network). The specific calculation methods are as follows: In the formula, is the tanh activation function; σ is the sigmoid function; respectively represent the weights between x t and the temporary memory cell z t , o t ; R z , R o respectively represent the recurrent weights between h t-1 and the temporary memory cell z t , o t ; b o are all bias terms. Based on the above information processing process, the cell update at time t is performed, and the specific calculation process is as follows: c t = f t c t-1 + i t z t , where c t-1 represents the cell state at time t - 1. For the forget gate and input gate using the exponential activation function, to control the scale of the information flow and help stabilize the learning process of the network, a normalized state is introduced, and the specific calculation is as follows: n t = f t n t-1 + i t , where n t-1 represents the normalized state at time t - 1. Finally, based on the cell state c t and the normalized state n t at the current time, the information finally input to the next time is determined, that is, the hidden state h t at time t, and the specific calculation method is as follows: h t = o t (c t / n t ); The structural schematic diagram of the sLSTM model is as Figure 5As shown, the data is embedded in the pre-LayerNorm residual layer and selectively passes through a causal convolutional layer with a window size of 4, which includes the Swish activation function for the input gate and the forget gate. Then, for all input gates, forget gates, output gates i, f, o, and the temporary memory cell z, the input data passes through a block diagonal linear layer with four diagonal blocks or "heads". These diagonal blocks correspond to the pre-activation of the recurrent gates of the previous hidden state. The main output of the sLSTM block is the updated hidden state. This hidden state is the result of being processed by the block diagonal linear layer and further processed by a GroupNorm layer. Then, the obtained hidden state is projected through a gate with a GeLU activation function and a projection factor of 4 / 3 to adjust the dimension of the hidden state to the required output dimension. Finally, an output sequence corresponding to the input sequence is output, but each element has been updated to a hidden state containing more context information. By introducing the exponential gating mechanism and the memory mixing mechanism into the LSTM (Long Short-Term Memory Network), the sLSTM model is obtained, breaking through the limitations of the output range and information sharing of the LSTM (Long Short-Term Memory Network). Furthermore, the sLSTM model is used to extract the features of the modal components to obtain the first time series feature vector, realizing the capture of the long-term dependence relationships existing in the data, solving the technical problems of being difficult to process non-linear data and difficult to capture the long-term dependence relationships existing in the data, and significantly improving the accuracy of electricity price prediction.

[0037] In this embodiment, the output of the exponential activation function is also related to the width of the historical window. By reasonably obtaining the width of the historical window, the output of the exponential activation function is restricted by the width. Since electricity price data usually exhibits significant daily, weekly, and monthly periodic patterns, these periodic characteristics are particularly prominent in adjacent dates or time periods. Therefore, in electricity price prediction, reasonably setting the width of the historical window and fully capturing and utilizing these periodic dependence relationships are crucial for improving the prediction performance. An overly short historical window may cause the model to fail to fully learn the historical laws of electricity price fluctuations, while an overly long historical window may introduce too much noise information, ignore the recent electricity price change trends, and interfere with the prediction effect of the model. Therefore, to comprehensively cover the daily, weekly, and monthly periodic characteristics of electricity price data and accurately analyze the electricity price cycle with the strongest correlation with the current electricity price, this paper selects multiples of 12 as the width of the historical window for prediction and comparative analysis according to the specific data characteristics and prediction requirements, and finally selects the optimal width of the historical window to achieve the best prediction performance. The comparison of evaluation indicators for different widths of the historical window is shown in Table 2: Table 2. Comparison Table of Evaluation Indicators for Different Widths of the Historical Window As can be seen from Table 2, as the width of the historical window gradually increases, the prediction performance of the model shows a trend of first increasing and then stabilizing. However, an overly long historical window may also lead to a decrease in prediction accuracy. Specifically, when the historical window is set to a relatively short 12 hours (i.e., half a day), the various error indicators of the model are relatively high, while the R 2 value is relatively low, indicating that the prediction accuracy at this time needs to be improved. When the historical window is extended to 24 hours (i.e., 1 day), the prediction performance of the model has been significantly improved. The MAE, RMSE, and MAPE all show a downward trend, and at the same time, the R 2 value increases significantly. This clearly reflects that increasing the length of historical data helps the model to learn more deeply the historical laws of electricity price fluctuations, thereby improving prediction accuracy. When the historical window reaches 72 hours (i.e., 3 days), the prediction performance of the model reaches a significant peak. The MAE and RMSE drop to 1.48 and 2.04 respectively, and the MAPE is as low as 0.06, while the R 2 value is as high as 0.97. This result fully demonstrates the importance of increasing the width of the historical window within a reasonable range for improving the accuracy of electricity price prediction. However, when the width of the historical window continues to increase, the performance of the model begins to show a downward trend, and the MAE, RMSE, and MAPE all rise to some extent, and the R 2 value also decreases. Especially when the width of the historical window reaches 336 hours (i.e., 14 days), the prediction performance of the model deteriorates further. This may be because an overly long historical window introduces too much noise information, interfering with the prediction effect of the model. To sum up, increasing the width of the historical window within a reasonable range helps to improve the accuracy of the electricity price prediction model. Under the research framework and dataset of this paper, the correlation between the current electricity price and the electricity prices in the past 72 hours is the strongest, which also indicates that a window width of 72 hours can reasonably balance the long-term dependence relationship and the capture of recent electricity price change trends, so as to achieve more accurate prediction. Therefore, a historical window width of 72 hours is the most suitable choice for electricity price prediction.

[0038] S3: Based on the first time series feature vector and the second historical electricity price data, use the Transformer model to obtain the second time series feature vector representing short-term dependence relationships, and obtain the short-term electricity price prediction result based on the first time series feature vector and the second time series feature vector.

[0039] The above S3 includes: S31: Map the feature sequences corresponding to the first time series feature vector and the second historical electricity price data to a high-dimensional space through the Transformer model to obtain a high-dimensional sequence, and add position encoding representing the time order to the high-dimensional sequence; S32: Obtain attention weights based on the position encoding, and obtain the second time series feature vector based on the attention weights.

[0040] The Transformer model is a deep learning model architecture mainly used to handle sequence-to-sequence tasks in natural language processing. Its core innovation is that through the self-attention mechanism, the model can consider all positions in the sequence simultaneously, and by processing the entire sequence in parallel, it greatly improves the computational efficiency. Considering that the sLSTM model has the ability to extract temporal features from historical electricity price data and capture long-term dependencies in time series, while the Transformer can capture short-term patterns in sequence data through the self-attention mechanism, and through multiple stacks of encoder and decoder layers, it can further fuse and process the features extracted by the sLSTM, and has the ability to capture complex data. The sLSTM-Transformer hybrid prediction model constructed in this paper effectively combines the advantages of both to improve the prediction accuracy of electricity price time series data. By inputting the first time series feature vector representing long-term dependencies and the second historical electricity price data into the Transformer model suitable for capturing short-term dependencies, the second time series feature vector representing short-term dependencies can be fully and accurately obtained, and then electricity price prediction is carried out by fully capturing long-term and short-term dependencies, further improving the accuracy of electricity price prediction.

[0041] In this embodiment, evaluation metrics such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are also used to comprehensively evaluate the prediction results to verify the prediction accuracy and generalization performance of the model. Specifically, the smaller the values of RMSE, MAPE, and MAE, the higher the prediction accuracy of the model; while R 2 is a key indicator to measure the prediction accuracy of the model, which reflects the goodness of fit between the predicted value of the model and the actual observed value. R 2 ranges from [0, 1], and the closer it is to 1, the better the fitting effect of the model. The mathematical expressions of the four evaluation metrics are as follows: where N represents the number of samples; y i represents the actual value of historical electricity price; represents the predicted value of electricity price; represents the average value of historical electricity price. To more deeply verify the advantages of the model proposed in this study, such as Figure 6As shown in the figure, the model of the present invention, VMD-CPO-sLSTM-Transformer (a hybrid model composed of an optimized modal decomposition model, an sLSTM model, and a Transformer model), is compared and analyzed with the CNN-LSTM model (a hybrid model composed of a convolutional neural network and a long short-term memory network), CNN-LSTM-Attention (a hybrid model composed of a convolutional neural network, a long short-term memory network, and an attention mechanism), and the Transformer model. The experimental results are as Figure 7 shown in Table 3: Table 3. Comparison Table of Evaluation Metrics for Different Prediction Models It can be Figure 7 seen that compared with other models, the predicted electricity price curve of the model in this paper is closer to the real electricity price curve, and it has stronger learning ability for historical electricity prices. As can be seen from Table 3, although the CNN-LSTM model provides certain prediction ability, there is still room for improvement in terms of error control and fitting degree. When the attention mechanism is introduced into CNN-LSTM, the prediction model is reduced to 2.79, 3.95, and 0.12 in terms of MAE, RMSE, and MAPE respectively, while R 2 is increased to 0.9, showing the effectiveness of the attention mechanism in improving prediction accuracy and model fitting degree. The MAE of the Transformer model is 3.33, slightly lower than that of CNN-LSTM, but the RMSE is reduced to 3.83, the MAPE is 0.15, and R 2 reaches 0.91. This indicates that the Transformer model has certain advantages in error control and model generalization ability, especially in dealing with the long-term dependence relationship of time series data. Finally, the R 2 of the model in this paper is as high as 0.97. Other indicators are also better than other models, fully demonstrating the excellent performance of the model in this paper in the electricity price prediction task. This model not only effectively combines the advantages of sLSTM in dealing with short-term dependencies of time series and Transformer in capturing long-term dependencies, but also further improves the prediction accuracy and generalization ability of the model through a composite structure.

[0042] Embodiment 2: This embodiment also provides a short-term electricity price prediction device based on crown porcupine optimization and deep learning, which is applicable to the short-term electricity price prediction method based on crown porcupine optimization and deep learning, and includes a modal component acquisition module, a dependence relationship acquisition module, and a short-term electricity price prediction result acquisition module; The modal component acquisition module is used to optimize the VMD model based on crown porcupine optimization to obtain an optimized VMD model, and decompose the first historical electricity price data into modal components based on the optimized VMD model; The dependency acquisition module is used to extract features from modal components based on the sLSTM model to obtain a first time series feature vector representing long-term dependencies, and is also used to obtain a second time series feature vector representing short-term dependencies based on the first time series feature vector and the second historical electricity price data using the Transformer model; The short-term electricity price prediction result acquisition module is used to obtain a short-term electricity price prediction result based on the first time series feature vector and the second time series feature vector. It further includes an sLSTM model acquisition module, which is used to introduce an exponential gating mechanism and a memory mixing mechanism into the LSTM to obtain the sLSTM model.

[0043] In this embodiment, introducing the exponential gating mechanism and the memory mixing mechanism into the LSTM (Long Short-Term Memory network) to obtain the sLSTM model is specifically as follows: In the input gate and the forget gate, the sLSTM introduces the exponential gating mechanism, and calculates the activation value of the gate through an exponential function, so that the output of the gate is no longer limited between 0 and 1, but can be dynamically adjusted as needed. This mechanism allows a wider range of information to pass through, making the information transfer more flexible and better capturing and utilizing the pattern of information flow; In the original LSTM (Long Short-Term Memory network), the update of the memory cell state is achieved through a simple linear combination, which may cause information to be gradually lost or blurred during the transfer process. The memory mixing mechanism promotes the complex interaction and information transfer between memory units within the same head at different time nodes, enabling it to more effectively capture the long-term dependency characteristics in time series data, thereby improving the performance and prediction accuracy of the model. The characteristic of processing nonlinear data through VMD realizes the processing of nonlinear data. At the same time, the VMD is optimized through the crown porcupine to avoid the problem of insufficient data decomposition. By introducing the exponential gating mechanism and the memory mixing mechanism into the LSTM to obtain the sLSTM model, the limitations of the output range and information sharing of the LSTM are broken through. Then, the sLSTM model is used to extract features from modal components to obtain the first time series feature vector, realizing the capture of the long-term dependencies existing in the data, solving the technical problems of difficult processing of nonlinear data and difficult capture of long-term dependencies existing in the data, and significantly improving the accuracy of electricity price prediction; At the same time, by inputting the first time series feature vector representing long-term dependencies and the second historical electricity price data into the Transformer model suitable for capturing short-term dependencies, the second time series feature vector representing short-term dependencies can be fully and accurately obtained. Then, through the fully captured long-term and short-term dependencies for electricity price prediction, the accuracy of electricity price prediction is further improved.

[0044] The above-described specific embodiments are the preferred embodiments of the short-term electricity price prediction method and device based on the crown porcupine optimization and deep learning of the present invention, and do not limit the specific implementation scope of the present invention. The scope of the present invention includes but is not limited to this specific embodiment. Any equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A short-term electricity price forecasting method based on crested porcupine optimization and deep learning, characterized in that: The following steps are involved: S1: optimizing the VMD model based on the crown porcupine optimization to obtain an optimized VMD model, and decomposing the first historical electricity price data into modal components based on the optimized VMD model; S2: Introduce the exponential gating mechanism and the memory hybrid mechanism into LSTM to obtain the sLSTM model. Based on the sLSTM model, extract the features of the modal components to obtain the first time series feature vector representing the long-term dependency relationship. S3: Based on the first time series feature vector and the second historical electricity price data, a Transformer model is used to obtain a second time series feature vector representing the short-term dependency relationship, and a short-term electricity price prediction result is obtained based on the first time series feature vector and the second time series feature vector.

2. The short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to claim 1 is characterized in that: In S1, the step of optimizing the VMD model based on the crown porcupine optimization to obtain the optimized VMD model includes: Randomly generate the position of each individual in the initial population, where the position information of each individual in the initial population represents a set of parameters to be screened for optimizing the VMD model; Calculate the fitness of each individual in the initial population and store the optimal solution of the population. In the exploration phase of the crown porcupine optimization, guide each individual in the initial population to move to the position corresponding to the optimal solution of the population to update the initial population range and obtain the first population range. In the development phase of the crown porcupine optimization, guide each individual in the first population range to move to the position corresponding to the optimal solution of the population to update the first population range. The final parameters are obtained according to the first population range and the population optimal solution, and the final parameters are used to replace the initial parameters in the VMD model to obtain the optimized VMD model.

3. The short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to claim 1 is characterized in that: In S1, decomposing the first historical electricity price data into modal components based on the optimized VMD model includes: S1a: Introduce the convergence parameters into the optimized VMD model to obtain the augmented Lagrangian expression, iteratively update the initial modal components based on the augmented Lagrangian expression, and judge whether the initial modal components meet the preset conditions. If they meet the preset conditions, the initial modal components are the modal components. Otherwise, update the convergence parameters based on the Wiener filter and execute S1a.

4. The short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to claim 1 is characterized in that: In S2, the exponential gating mechanism and the memory hybrid mechanism are introduced into the LSTM to obtain the sLSTM model, including: The exponential gating mechanism is introduced into the activation value calculation mechanism of LSTM, and the memory mixing mechanism is introduced into the cell state update mechanism of LSTM to obtain the sLSTM model.

5. The short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to claim 1 is characterized in that: In S2, the feature extraction of the modal component based on the sLSTM model to obtain the first time series feature vector representing the long-term dependency relationship includes: The output of the exponential activation function in the sLSTM model is used to remove old information of the cell unit in the sLSTM model, and the output of the exponential activation function is used to control the new information stored in the cell unit, thereby updating the sLSTM model; Perform a logarithmic operation on the output of the updated sLSTM model, obtain the rated state and the normalized state according to the logarithmic operation result, introduce the rated state and the normalized state into the updated sLSTM model, and perform feature extraction on the modal component through the updated sLSTM model to obtain the first time series feature vector.

6. The short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to claim 5 is characterized in that: The method of extracting features from the modal component based on the sLSTM model to obtain a first time series feature vector representing the long-term dependency relationship also includes: The initial output of the exponential activation function is obtained based on the modal component characteristics, and based on the initial output, the process prediction result is obtained using the third historical electricity price data. The initial output is updated based on the process prediction result and the actual prediction result to obtain the output of the exponential activation function.

7. The short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to claim 5 is characterized in that: The method of using the output of the exponential activation function in the sLSTM model to remove old information of the cell units in the sLSTM model includes: f=exp(Wx t +Rh t-1 +b); Where f is the output of the forget gate in the sLSTM model; x t is the input value of the forget gate at time t; W is x t The weight vector between and f; h t-1 is the hidden state at time t-1; R is h t-1 The recursive weight between and f, and b is the bias term.

8. The short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to claim 1 is characterized in that: The S3 includes: S31: Mapping the first time series feature vector and the feature sequence corresponding to the second historical electricity price data to a high-dimensional space through a Transformer model to obtain a high-dimensional sequence, and adding a position code representing the time order to the high-dimensional sequence; S32: Obtain an attention weight based on the position encoding, and obtain a second time series feature vector based on the attention weight.

9. A short-term electricity price forecasting device based on crested porcupine optimization and deep learning, applicable to the short-term electricity price forecasting method based on crested porcupine optimization and deep learning according to any one of claims 1 to 8, characterized in that: It includes a modal component acquisition module, a dependency acquisition module and a short-term electricity price forecast result acquisition module; The modal component acquisition module is used to optimize the VMD model based on the crown porcupine optimization to obtain the optimized VMD model, and decompose the first historical electricity price data into modal components based on the optimized VMD model; The dependency acquisition module is used to extract features from modal components based on the sLSTM model to obtain a first time series feature vector representing a long-term dependency, and is also used to obtain a second time series feature vector representing a short-term dependency using a Transformer model based on the first time series feature vector and the second historical electricity price data; The short-term electricity price forecast result acquisition module is used to acquire the short-term electricity price forecast result based on the first time series feature vector and the second time series feature vector.

10. The short-term electricity price prediction device based on crested porcupine optimization and deep learning according to claim 9, characterized in that: It also includes an sLSTM model acquisition module, which is used to introduce the exponential gating mechanism and the memory hybrid mechanism into the LSTM to obtain the sLSTM model.

Images