A short-term load forecasting method and system for power spot market across seasons

Abstract

This invention relates to a method and system for short-term load forecasting across seasons in the electricity spot market. The method first collects electricity load, meteorological, and temporal characteristic data to construct a multi-dimensional input feature set, which is then divided into a training set and a validation set. Next, variational mode decomposition is used to decompose the original load data into three modal components, corresponding to three types of input feature sets. Then, with the goal of maximizing the validation set determination coefficient, seasonally differentiated hyperparameter optimization is performed using a particle swarm optimization algorithm. Temporal dependency prediction sub-models and multi-feature association prediction sub-models are constructed based on the optimized Long Short-Term Memory Network and Random Forest, respectively. The three types of input feature sets are then input into their respective sub-models, outputting two types of prediction results. Finally, the two types of results are weighted and integrated to obtain the short-term load forecast result. Compared with existing technologies, this invention has advantages such as significantly improved prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a method and system for short-term load forecasting across seasons for the electricity spot market. Background Technology

[0002] Currently, the power system is gradually transitioning from traditional dispatching to market-based operation. As the core platform for optimizing energy resource allocation, the efficient operation of the electricity spot market heavily relies on accurate short-term load forecasting. Short-term load forecasting is a core technical support for ensuring the safe and economical operation of the power system and improving the efficiency of spot market transactions: On the generation side, load forecasting is necessary to formulate unit combination and bidding strategies to avoid redundant reserves or a surge in peak-valley regulation costs; on the grid side, load forecasting is needed to optimize transmission channel utilization and reduce congestion risks and network losses; and on the user side, load forecasting allows participation in demand response, leveraging peak-valley price differences to reduce electricity costs.

[0003] However, short-term load forecasting in the electricity spot market environment faces multiple technical challenges: Load data has complex characteristics: affected by factors such as the intermittency of renewable energy (fluctuations in wind power and photovoltaic output) and adjustments in user electricity consumption behavior (load transfer guided by electricity prices), the load sequence exhibits strong nonlinearity and nonstationarity. Traditional prediction methods based on statistical models (such as ARIMA and exponential smoothing) or single machine learning models (such as independent LSTM and RF) are difficult to capture multi-scale fluctuation patterns, and the prediction accuracy is limited. Multi-factor coupled interference: The load is not only affected by meteorological conditions (high temperature cooling in summer, low temperature heating in winter, short-term weather events in spring and autumn) and time characteristics (differences in electricity consumption on weekdays / holidays, and intraday peak-valley cycles), but also forms a two-way feedback with the electricity spot market price signal (users reduce unnecessary loads during periods of high electricity prices). Existing methods mostly consider a single factor in isolation and have not achieved multi-factor collaborative modeling. Insufficient seasonal adaptability: The load fluctuation patterns differ significantly across seasons (sustained high load in summer, peak heating season in winter, and stable fluctuations in spring and autumn). Traditional models, which use uniform parameters or fixed structures, cannot adapt to the seasonal differences, resulting in unstable cross-seasonal forecasting performance and making it difficult to meet the needs of the electricity spot market for continuous operation throughout the year.

[0004] To address these challenges, researchers have attempted to introduce modal decomposition (such as EMD and EEMD) and multi-model fusion methods. However, traditional decomposition methods are prone to modal aliasing, and model parameters often rely on empirical settings without being specifically optimized to take into account seasonal characteristics and market factors. Therefore, they still cannot effectively balance prediction accuracy, stability, and market adaptability. For example, the invention patent with publication number CN119448265A discloses a power load forecasting method that integrates Transformer and RF. Although this method introduces K-Shape clustering and STL decomposition to process load data, uses Transformer to extract features, and combines random forest for regression prediction, thus improving prediction performance to some extent, it still has the following shortcomings: Limitations of the decomposition method: The STL (Seasonal-Trend decomposition using Loess) decomposition method used in this method is based on statistical principles. For spot market load data with strong nonlinearity and high-frequency random noise, its adaptive ability is weaker than that of variational mode decomposition. It is difficult to completely separate random fluctuation components from periodic components, which can easily lead to impure feature extraction. Lack of market factor integration: The scheme only considers historical load and meteorological data, ignoring the guiding role of electricity price signals, which are the most critical factors in the electricity spot market, on load. This leads to distorted forecasts during periods of sharp price fluctuations (such as load reductions caused by price responses). Model targeting and optimization deficiencies: This method does not match differentiated model structures (such as using only RF regression) for different component characteristics after decomposition (trend term vs. fluctuation term), and lacks a global optimization mechanism for all model hyperparameters (such as PSO), making it difficult to automatically adapt the optimal parameter combination in different seasonal scenarios, thus limiting the robustness of the model in complex and volatile market environments.

[0005] In summary, current short-term load forecasting in the electricity spot market suffers from several problems, including difficulty in accurately separating multi-scale coupling characteristics, neglecting the price signal guidance mechanism, and the inability of model parameters to adaptively adjust with seasonal characteristics, resulting in insufficient forecast accuracy and stability. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and system for short-term load forecasting across seasons in the electricity spot market.

[0007] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a method for short-term load forecasting across seasons in the electricity spot market is provided, characterized in that the method steps include: S1. Collect raw power load data and corresponding meteorological and temporal feature data for the same period, construct a multi-dimensional input feature set, and build a training set and a validation set based on the feature set; S2. Using the variational mode decomposition algorithm, the original power load data in the multi-dimensional input feature set is decomposed into three types of mode components: random fluctuation component, intra-day periodic component, and seasonal trend component. Based on the correspondence of the three types of mode components, the multi-dimensional input feature set is divided into random fluctuation input feature set, intra-day periodic input feature set, and seasonal trend input feature set. S3. Using the particle swarm optimization algorithm, with the goal of maximizing the coefficient of determination on the validation set, the hyperparameters of the Long Short-Term Memory Network and the Random Forest are seasonally differentiated and optimized based on the training set. The optimized Long Short-Term Memory Network is used to construct a time-dependent prediction sub-model, and the optimized Random Forest is used to construct a multi-feature association prediction sub-model. S4. Input the seasonal trend input feature set into the time-series dependent prediction sub-model and output the first prediction result. Input the random fluctuation input feature set and the intraday cycle input feature set into the multi-feature association prediction sub-model and output the second prediction result. S5. The first and second forecast results are weighted and integrated to obtain the short-term load forecast result.

[0008] As a preferred technical solution, when constructing the multi-dimensional input feature set in S1, the process also includes handling missing data values. Specifically, the collaborative linear interpolation method is used to fill in the missing values ​​of the original power load data. The specific process of the collaborative linear interpolation method includes: The interpolation base value is calculated based on the effective load values ​​adjacent to the missing time in the original power load data. By introducing a preset electricity price correction factor, and combining the average electricity price of adjacent time periods, the electricity price of the missing time period, and the preset electricity price correction factor, the interpolation base value is corrected to obtain the corrected missing load value. The missing load values ​​are filled in to the missing moments of the original power load data to form a complete load data sequence; The 3σ criterion is used to detect the complete load data sequence. The mean μ and standard deviation σ of the data sequence are calculated, and data that exceed the interval [μ-3σ, μ+3σ] are identified as outliers and removed.

[0009] As a preferred technical solution, the parameter configuration of the variational mode decomposition algorithm includes: Set the number of decomposed modes K=3, the penalty factor α=2000, the noise tolerance τ=0, and do not force the decomposition of DC components; When using the variational mode decomposition algorithm to decompose the original power load data into three types of modal components, the original power load data is adaptively separated according to the frequency from high to low or from low to high. Among them, the random fluctuation component corresponds to the highest frequency mode; the intra-day periodic component corresponds to the intermediate frequency mode; and the seasonal trend component corresponds to the lowest frequency mode.

[0010] As a preferred technical solution, in S3, the hyperparameters of the Long Short-Term Memory Network specifically include the number of hidden layer neurons, the number of stacked layers, and the learning rate; the hyperparameters of the Random Forest specifically include the number of decision trees and the maximum depth.

[0011] As a preferred technical solution, in S3, when using the particle swarm optimization algorithm for seasonal difference optimization, the specific configuration includes: The particle swarm optimization algorithm is set to have 20 particles and a maximum number of iterations of 40. When optimizing the Long Short-Term Memory network, the optimization search range for the number of hidden layer neurons is set to [32, 256], the optimization search range for the number of stacked layers is set to [1, 5], and the optimization search range for the learning rate is set to [0.0001, 0.01]. When optimizing the Random Forest, the optimization search range for the number of decision trees is set to [50, 150], and the optimization search range for the maximum depth is set to [3, 25].

[0012] As a preferred technical solution, the seasonal differentiation optimization in S3 satisfies the following seasonal adaptation rules: In the summer scenario, the learning rate of the Long Short-Term Memory network obtained by the particle swarm optimization algorithm is set to the interval [0.0001, 0.01), and the maximum depth of the random forest is set to the interval [19, 20]. In spring or autumn scenarios, the number of stacked layers of the long short-term memory network obtained by the particle swarm optimization algorithm is set to the interval [2, 4] and the learning rate is set to 0.01. The maximum depth of the random forest is set to 17. In the winter scenario, the number of stacked layers of the Long Short-Term Memory network obtained by the Particle Swarm Optimization algorithm is set to the interval [3, 4], and the maximum depth of the Random Forest is set to the interval [19, 20].

[0013] As a preferred technical solution, the specific prediction logic of each sub-model in S4 is as follows: For the time-dependent prediction sub-model, the seasonal trend input feature set is used as input. The gating mechanism of the long short-term memory network is used to extract the long-term memory features and short-term evolution rules of the load trend sequence in the time dimension. Based on the long-term memory features and short-term evolution rules, the load trend at the next moment is extrapolated in time, thereby outputting the first prediction result. For the multi-feature association prediction sub-model, the random fluctuation input feature set and the intraday cycle input feature set are used as inputs. The ensemble decision tree structure of random forest is used to mine the nonlinear coupling relationship between meteorological data and time feature data and load fluctuation components. Based on the coupling relationship, the load fluctuation value at the next moment is regressed and predicted, thus outputting the second prediction result.

[0014] As a preferred technical solution, the specific process of weighted integration of the first prediction result and the second prediction result in S5 includes: Using validation set data, the mean absolute percentage error of the time-dependent prediction sub-model and the multi-feature association prediction sub-model were calculated respectively. The prediction accuracy of each sub-model is evaluated based on the mean absolute percentage error, and the weighting coefficient of the prediction results of each sub-model is determined according to the prediction accuracy of each sub-model. Based on the weighting coefficients, the first and second forecast results are weighted and summed to obtain the short-term load forecast result.

[0015] According to another aspect of the present invention, a short-term load forecasting system for the electricity spot market across seasons is provided, the system comprising: The data acquisition and processing module is configured to acquire raw power load data and meteorological data and time feature data for the corresponding time period, construct a multi-dimensional input feature set, and construct a training set and a validation set based on the multi-dimensional input feature set; The modal decomposition and feature allocation module is configured to use the variational modal decomposition algorithm to decompose the original power load data in the multi-dimensional input feature set into three types of modal components: random fluctuation component, intra-day periodic component, and seasonal trend component. Based on the three types of modal components, the multi-dimensional input feature set is divided into random fluctuation input feature set, intra-day periodic input feature set, and seasonal trend input feature set. The model optimization and construction module is configured to use the particle swarm optimization algorithm to maximize the coefficient of determination on the validation set. Based on the training set, the hyperparameters of the Long Short-Term Memory Network and the Random Forest are optimized seasonally. The optimized Long Short-Term Memory Network is used to construct a temporal dependency prediction sub-model, and the optimized Random Forest is used to construct a multi-feature association prediction sub-model. The multi-component prediction module is configured to input the seasonal trend input feature set into the time-series dependent prediction sub-model and output the first prediction result, and input the random fluctuation input feature set and the intraday cycle input feature set into the multi-feature association prediction sub-model and output the second prediction result. The results integration module is configured to weight and integrate the first and second forecast results to obtain the short-term load forecast results.

[0016] As a preferred technical solution, the model optimization and construction module includes a pre-built seasonal scene recognition unit: The seasonal scene recognition unit is configured as follows: Extract month information for the date to be predicted from time feature data; Based on the monthly information, the current period to be predicted is determined to be a summer scenario, a winter scenario, or a spring / autumn scenario; Based on the determined scenario type, the corresponding long short-term memory network hyperparameter search interval and random forest hyperparameter search interval are retrieved from the system database, and the search interval is passed as a constraint condition to the particle swarm optimization algorithm.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs variational mode decomposition (VMD) to replace traditional decomposition methods, accurately decomposing the original load data into three components: random fluctuations, intraday cycles, and seasonal trends. A particle swarm optimization (PSO) algorithm is then used to automatically optimize the hyperparameters of the LSTM and RF models, aiming to maximize the validation set determination coefficients. This technical solution effectively overcomes the mode aliasing problem common in traditional decomposition methods, as well as the performance bottleneck caused by the reliance on empirically set model parameters and the lack of targeted optimization. Addressing the multi-factor coupling characteristics of the electricity spot market load, VMD multi-scale decomposition separates the core load features. Combined with PSO-optimized LSTM and RF sub-models for collaborative prediction, the cross-seasonal prediction determination coefficients remain above 0.95, with an average absolute percentage error below 2%, significantly improving prediction accuracy compared to traditional single-model predictions. This allows for precise capture of the load fluctuation patterns in the spot market.

[0018] 2. In this invention, seasonal differentiation optimization satisfies the seasonal adaptation rules. The system has a pre-set seasonal scene recognition unit that automatically matches seasonal scenes. Differentiated hyperparameter search intervals are set for model optimization. Combined with the physical constraints of VMD parameters and the error weighted integration mechanism, this method and the corresponding system can maintain stable performance in different scenarios such as high load fluctuations in summer, continuous peaks in winter, and stable conditions in spring and autumn. This avoids seasonal adaptation failure caused by a single parameter. At the same time, the weighted integration strategy further reduces the impact of random disturbances. The cross-seasonal prediction accuracy fluctuation is less than 0.5%, which shows strong seasonal adaptability, excellent robustness, and stable prediction throughout the year.

[0019] 3. This invention addresses the problem of ineffective integration of multiple influencing factors in the prior art by incorporating electricity price signals into the data preprocessing stage using collaborative linear interpolation to fill in missing values, and by employing heterogeneous models (LSTM predicts trends, RF predicts fluctuations) for different components in the prediction stage. By introducing an electricity price correction coefficient and integrating the electricity price signal into the data preprocessing, the invention fully considers the demand response mechanism of price-driven load in the spot market, repairs feature gaps caused by missing data, and improves the economic accuracy of the data. Through the specific prediction logic of each sub-model, LSTM is used to extract long-term memory features, and random forests are used to mine the nonlinear coupling relationship between meteorological and temporal features and load fluctuations, achieving accurate extraction of multi-scale load features.

[0020] 4. In this invention, a short-term load forecasting system is constructed that includes the entire process of data acquisition, decomposition, optimization, prediction and integration.

[0021] This technical solution can be directly applied to core scenarios in the electricity spot market, such as generation-side bidding, grid-side dispatching, and user-side demand response. It can also be extended to the field of multi-energy load forecasting, such as cold load and hot load, providing reliable technical support for the efficient operation of the spot market. It has broad system application value and is highly practical. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the short-term load forecasting method for the electricity spot market across seasons in this invention; Figure 2 This is a time series diagram of annual power load in the example; Figure 3a This is a schematic diagram of the waveform of the original signal from the spring test set load variational mode decomposition in the embodiment. Figure 3b This is a schematic diagram of the waveform of the main frequency of component 1 obtained by decomposition in spring in the example. Figure 3c This is a schematic diagram of the second frequency waveform of the component obtained by decomposition in spring in the example. Figure 3d This is a schematic diagram of the micro-frequency waveform obtained by decomposition in the spring in the example. Figure 4a This is a waveform diagram of the original signal from the variational mode decomposition of the test set load in summer, as shown in the example. Figure 4b This is a waveform diagram of the main frequency of component 1 obtained by decomposition in summer in the example. Figure 4c This is a schematic diagram of the second frequency waveform of the component obtained by decomposition in summer in the example. Figure 4dThis is a schematic diagram of the waveform of the 3 micro-frequency component obtained by decomposition in summer in the example. Figure 5a This is a waveform diagram of the original signal from the load variational mode decomposition of the test set in autumn in the example. Figure 5b This is a waveform diagram of the main frequency of component 1 obtained by decomposition in autumn in the example. Figure 5c This is a schematic diagram of the second frequency waveform of the component obtained by decomposition in autumn in the example. Figure 5d This is a schematic diagram of the waveform of the 3 micro-frequency component obtained by decomposition in autumn in the example. Figure 6a This is a waveform diagram of the original signal from the variational mode decomposition of the load in the test set during winter, as shown in the example. Figure 6b This is a waveform diagram of the main frequency of component 1 obtained by decomposition in winter in the embodiment. Figure 6c This is a schematic diagram of the second frequency waveform of the components obtained from the decomposition in winter in the example. Figure 6d This is a schematic diagram of the waveform of the 3 micro-frequency component obtained by decomposition in winter in the embodiment. Figure 7a This is a schematic diagram comparing the prediction results of spring component 1 in the embodiment; Figure 7b This is a schematic diagram comparing the prediction results of spring component 2 in the embodiment; Figure 7c This is a schematic diagram comparing the prediction results of spring component 3 in the example. Figure 8a This is a schematic diagram comparing the prediction results of summer component 1 in the embodiment; Figure 8b This is a schematic diagram comparing the prediction results of summer component 2 in the embodiment; Figure 8c This is a schematic diagram comparing the prediction results of summer component 3 in the embodiment; Figure 9a This is a schematic diagram comparing the prediction results of autumn component 1 in the embodiment. Figure 9b This is a schematic diagram comparing the prediction results of autumn component 2 in the embodiment; Figure 9c This is a schematic diagram comparing the prediction results of autumn component 3 in the example. Figure 10a This is a schematic diagram comparing the prediction results of winter component 1 in the embodiment. Figure 10b This is a schematic diagram comparing the prediction results of winter component 2 in the embodiment; Figure 10c This is a schematic diagram comparing the prediction results of winter component 3 in the embodiment; Figure 11a This is a comparison chart of load forecast fitting in spring in the examples; Figure 11b This is a comparison chart of summer load forecast fitting in the examples; Figure 11c This is a comparison chart of load forecast fitting in autumn in the example; Figure 11d This is a comparison chart of load forecast fitting in winter in the example; Figure 12 This is a schematic diagram of the error distribution of different models in the embodiment. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] This solution is mainly used to address the problems of existing forecasting models being unable to cope with the strong nonlinearity and non-stationarity of load data in the electricity spot market environment, as well as the inability to effectively integrate the coupled influence of multiple factors such as meteorological conditions, electricity price signals, and seasonal patterns, resulting in low forecasting accuracy and poor market adaptability.

[0025] Specifically, the short-term load forecasting process of this scheme is as follows: Based on the half-hour interval power load data of electricity spot market users, combined with meteorological data (temperature, humidity, wind speed, solar radiation) and time characteristic data (hour, weekday, holiday) for the corresponding time period, a multi-dimensional input feature set is constructed; Variational Mode Decomposition (VMD) is used to adaptively decompose the original load sequence into three types of modal components: high-frequency random fluctuation, medium-frequency intraday cycle, and low-frequency seasonal trend, avoiding the mode aliasing problem of traditional decomposition methods; Particle Swarm Optimization (PSO) algorithm is used to obtain the validation set coefficient of determination (R²). 2With the goal of maximizing performance, seasonally differentiated hyperparameter optimization is performed on the number of hidden layer neurons, stacking layers, and learning rate of the Long Short-Term Memory Network (LSTM), and the number of decision trees and maximum depth of the Random Forest (RF). For the feature differences of different frequency modal components, LSTM is used to capture temporal dependencies, and RF is used to mine multi-feature association patterns, constructing sub-prediction models. The prediction results of each sub-model are weighted and integrated to obtain the final short-term load forecast. This scheme can accurately capture the multi-scale fluctuation characteristics of load in the electricity spot market environment. Its cross-seasonal prediction performance is superior to traditional single models and basic hybrid models. It can achieve accurate extraction of multi-scale load features and in-depth mining of temporal patterns, ultimately achieving high-precision and robust prediction of short-term load in the electricity spot market, balancing prediction accuracy, market adaptability, and cross-seasonal stability. It provides reliable technical support for the efficient operation of the electricity spot market, the formulation of scenario-based prediction strategies, and transaction risk management. The specific implementation of this scheme is further described in detail below through specific examples.

[0026] Example 1 In this embodiment, a short-term load forecasting method for the electricity spot market across seasons is adopted, and the method flow is as follows: Figure 1 As shown, the specific steps include: S1. Collect raw power load data and corresponding meteorological and temporal feature data for the same period, construct a multi-dimensional input feature set, and build a training set and a validation set based on the feature set; S2. Using the variational mode decomposition algorithm, the original power load data in the multi-dimensional input feature set is decomposed into three types of mode components: random fluctuation component, intra-day periodic component, and seasonal trend component. Based on the correspondence of the three types of mode components, the multi-dimensional input feature set is divided into random fluctuation input feature set, intra-day periodic input feature set, and seasonal trend input feature set. S3. Using the particle swarm optimization algorithm, with the goal of maximizing the coefficient of determination on the validation set, the hyperparameters of the Long Short-Term Memory Network and the Random Forest are seasonally differentiated and optimized based on the training set. The optimized Long Short-Term Memory Network is used to construct a time-dependent prediction sub-model, and the optimized Random Forest is used to construct a multi-feature association prediction sub-model. S4. Input the seasonal trend input feature set into the time-series dependent prediction sub-model and output the first prediction result. Input the random fluctuation input feature set and the intraday cycle input feature set into the multi-feature association prediction sub-model and output the second prediction result. S5. The first and second forecast results are weighted and integrated to obtain the short-term load forecast result.

[0027] When constructing a multi-dimensional input feature set: Based on the raw electricity load data of users in the electricity spot market at half-hour intervals, combined with meteorological data and time feature data for the corresponding time periods, a multi-dimensional input feature set is constructed after processing missing data values ​​using the collaborative linear interpolation method. The raw electricity load data is half-hour interval sampling data covering at least one year to ensure that it contains complete seasonal features. The meteorological data includes temperature, humidity, wind speed and solar radiation, and the sampling frequency is consistent with that of the raw electricity load data. The time feature data includes hour markers (0-23), weekday markers (1-7) and binary markers for holidays (1 for holidays, 0 for weekdays). The linear interpolation method combines the load values ​​at adjacent times with the electricity spot market price signals for the corresponding time period to supplement the missing values.

[0028] When VMD adaptively decomposes the original load sequence: Variational mode decomposition was used to adaptively decompose the preprocessed raw power load data, resulting in three types of mode components: high-frequency random fluctuation component, medium-frequency intraday periodic component, and low-frequency seasonal trend component. The VMD parameter configuration is as follows: number of decomposition modes =3. Penalty Factor =2000, Noise Tolerance =0, no forced decomposition of DC component (DC=0), center frequency initialization mode is 1, convergence threshold tol=1×10 -7 ; The high-frequency random fluctuation component accounts for 10%-15% of the energy, mainly reflecting load fluctuations caused by short-term weather disturbances and random electricity consumption behavior of users; the medium-frequency intraday periodic component accounts for 20%-30% of the energy, mainly reflecting intraday peak and valley electricity consumption patterns; and the low-frequency seasonal trend component accounts for 55%-70% of the energy, mainly reflecting long-term load trends caused by seasonal meteorological changes.

[0029] When optimizing PSO seasonally differentiated hyperparameters: The PSO algorithm is used to determine the validation set coefficient of determination (R²). 2 With the goal of maximizing, seasonally differentiated optimization is performed on the hyperparameters of LSTM and RF; The core parameters of the PSO algorithm are configured as follows: number of particles 20, maximum number of iterations 40, inertia weight 0.9, and cognitive coefficient and social coefficient both 2.0; The hyperparameter optimization range for LSTM is: 32-256 hidden layer neurons, 1-5 stacked layers, and a learning rate of 0.0001-0.01. The hyperparameter optimization range for RF is: number of decision trees 50-150, maximum depth 3-25; The seasonal optimization is as follows: In the summer scenario, the LSTM learning rate is 0.0001-0.01 and the maximum RF depth is 19-20; in the spring / autumn scenario, the LSTM learning rate is 0.01, the number of stacked layers is 2-4, and the maximum RF depth is 17; in the winter scenario, the number of LSTM stacked layers is 3-4 and the maximum RF depth is 19-20.

[0030] When the sub-model makes predictions: To address the characteristic differences among high-frequency random fluctuation components, medium-frequency intraday cycle components, and low-frequency seasonal trend components, optimized LSTM is used to construct a time-dependent prediction sub-model, and optimized RF is used to construct a multi-feature association prediction sub-model, respectively, to obtain the prediction results for each modal component. The LSTM time-dependent prediction sub-model takes historical data of low-frequency seasonal trend components and corresponding meteorological and temporal characteristics as input. It uses forget gate, input gate and output gate to coordinate the update of memory cell state, capture long-term and short-term time-dependent relationships, and output the predicted value of low-frequency components. The RF multi-feature association prediction sub-model takes historical data of high / medium frequency components and a complete multi-dimensional feature set as input, and generates multiple decision trees through Bootstrap resampling. The coefficients determine the decision tree split point and output the predicted values ​​of high / medium frequency components. The formula for calculating the coefficient is: ,in For the node sample set, For the sample to belong to the first The probability of a class.

[0031] During weighted integration: normalization is performed based on the mean absolute percentage error (MAPE) of each sub-model, and after determining the weight coefficients, the prediction results of each modal component are weighted and integrated to obtain the final short-term load forecast value. Let the MAPEs of the sub-models corresponding to the three types of components be MAPE1, MAPE2, and MAPE3, then: ; ; ; Final short-term load forecast = ×Low-frequency component prediction value × High-frequency component prediction value ×Predicted value of mid-frequency component.

[0032] The specific implementation process of this method is as follows: I. Implementation Scenarios and Data Preparation This embodiment takes a comprehensive user in a city of a certain province as the research object. This user is a pilot participant in the province's electricity spot market. The load is affected by multiple factors such as seasonal weather (high temperature cooling in summer and low temperature heating in winter), electricity price signals (peak-valley time-of-use electricity price in the spot market), and industry production scheduling (three-shift production in industry). It has typical electricity spot market load characteristics, which can fully verify the practicality and superiority of this technical solution.

[0033] 1. Data Sources and Specifications All data used in this embodiment are actual collected data, specifically including: Raw power load data: Annual power load time series graph as shown below Figure 2 As shown, the half-hourly load data covers the period from July 1, 2024 to June 30, 2025, with a total of 17,520 records. The load unit is kW. The data is collected from smart meters on the user side and shows the load fluctuation characteristics of the four seasons: the peak load in winter is 350kW, the peak load in summer is 320kW, and the load in spring and autumn is stable in the range of 250-300kW. Meteorological data: The city's meteorological station released half-hourly data during the same period, including temperature, humidity, wind speed, and solar radiation, totaling 17,520 records; Time feature data: hour identifier, weekday identifier, and holiday binary identifier, generated based on calendar information and user production plan; Electricity spot market price data: The electricity spot market price data of the city during the same period, with half-hour intervals, in yuan / kWh, totaling 17,520 records, used for price correction during the data preprocessing stage.

[0034] 2. Data partitioning rules To verify the cross-seasonal prediction performance, the training set, validation set, and test set were divided according to season, as follows: Training set: This constitutes the main part of the total data volume and covers non-extreme working condition data for each season: spring (March-May), summer (June-August), autumn (September-November), and winter (December-February). It is used for basic model training. Validation set: One week of data randomly selected from each season (4 weeks in total) is used for hyperparameter optimization and model overfitting control in particle swarm optimization (PSO); Test set: In summer, the extreme high temperature week and one random week are selected; in winter, the extreme low temperature week and one random week are selected; and in spring and autumn, two random weeks are selected each for the final prediction accuracy verification.

[0035] II. Specific Implementation Steps This embodiment strictly follows the process of "constructing feature set → variational mode decomposition (VMD) → PSO optimization → sub-model prediction → weighted ensemble" in this scheme. The detailed operation and parameter settings for each step are as follows: Step 1: Construct a multi-dimensional input feature set 1.1 Data Preprocessing (Handling Missing and Outlier Values) The missing values ​​in the original load data and meteorological data were processed using the cooperative linear interpolation method (a total of 43 missing data entries were repaired, with reasons including meter communication interruption and meteorological station equipment maintenance). The specific operation is as follows: First, the baseline interpolation result is calculated based on the effective load values ​​at adjacent time points. The formula is: Interpolation baseline value = Load value at the previous effective time point. (Load value at the next valid time - Load value at the previous valid time) / (Load value at the next valid time - Load value at the previous valid time) (Missing time step - previous valid time step); Secondly, an electricity price adjustment coefficient is introduced. =1.1 (determined based on the fluctuation range of the city's spot market electricity price), adjusted by combining the average electricity price of adjacent periods and the electricity price of the missing period, the formula is: Corrected missing load value = Interpolation base value [1 [(Average electricity price in adjacent time periods - electricity price in missing time periods) / average electricity price in adjacent time periods], ensuring that the interpolation result reflects the guiding role of electricity price on users' electricity consumption behavior; Finally, using 3 The criteria remove outliers (28 in total, such as sudden load increases / decreases caused by industrial equipment failures) to ensure data validity.

[0036] 1.2 Feature Fusion The preprocessed raw load data, meteorological data, time characteristic data, and electricity price data are aligned according to the time dimension and spliced ​​together to form a multi-dimensional input feature set. The feature set is uniformly formatted as a CSV file for easy calling by subsequent modules.

[0037] Step 2: VMD Adaptive Decomposition of the Original Load Sequence The VMD decomposition is implemented using the vmdpy toolkit in Python. The parameters are configured as follows: Core parameter: Number of decomposed modes =3. Penalty Factor =2000, Noise Tolerance =0, no forced decomposition of DC component (DC=0), center frequency initialization mode=1, convergence threshold tol=1×10 -7 ; Decomposition process: The "original load data column" in the multi-dimensional input feature set constructed in step 1 is extracted separately and used as the VMD input sequence. After decomposition, three types of modal components are output, as shown in Figures 3, 4, 5 and 6, which correspond to the VMD decomposition results of power load in spring, summer, autumn and winter, respectively.

[0038] Figure 3a A schematic diagram of the waveform of the original signal from the variational mode decomposition of the load on the test set in spring; Figure 3b The waveform diagram of the main frequency of component 1 obtained from the spring decomposition is shown, with a center frequency of 0.021 and an energy ratio of 89.5%. Figure 3c The diagram shows the waveforms of the second frequency components obtained from the spring decomposition, with a center frequency of 0.103 and an energy percentage of 10.3%. Figure 3d This is a schematic diagram of the waveform of the micro-frequency obtained from the decomposition in spring, with a center frequency of 0.5 and an energy percentage of 0.2%.

[0039] Figure 4a A waveform diagram of the original signal from the variational mode decomposition of the load on the summer test set; Figure 4b The waveform diagram of the main frequency of component 1 obtained from the decomposition in summer is shown, with a center frequency of 0.027 and an energy percentage of 85.1%. Figure 4c The waveform diagram of the second frequency of the components obtained from the decomposition in summer is shown, with a center frequency of 0.025 and an energy percentage of 14.3%. Figure 4b The waveform diagram of the 3 micro-frequency component obtained from the decomposition in summer is shown, with a center frequency of 0.024 and an energy percentage of 0.6%.

[0040] Figure 5a A waveform diagram of the original signal from the variational mode decomposition of the load for the autumn test set; Figure 5b The waveform diagram of the main frequency of component 1 obtained from the decomposition in autumn is shown, with a center frequency of 0.027 and an energy percentage of 70.0%. Figure 5c The waveform diagram of the second frequency of the component obtained from the decomposition in autumn is shown, with a center frequency of 0.025 and an energy percentage of 29.7%. Figure 5d The waveform diagram of the 3 micro-frequency component obtained from the decomposition in autumn is shown, with a center frequency of 0.024 and an energy percentage of 0.3%.

[0041] Figure 6a For the winter test set load variational mode decomposition - raw signal; Figure 6b The component obtained from the decomposition in winter has a primary frequency of 1, a center frequency of 0.027, and an energy percentage of 63.4%. Figure 6c The second frequency of the component obtained from the decomposition in winter is 0.025, with a center frequency of 0.025 and an energy percentage of 35.9%. Figure 6d The component obtained from the decomposition in winter has a frequency of 3 micro-frequency, a center frequency of 0.024, and an energy percentage of 0.7%.

[0042] Figures 4a-6d The characteristics of each component in the spring / summer / autumn / winter season decomposition effect are as follows: High-frequency random fluctuation component: accounting for 10%-15% of energy, mainly reflecting load fluctuations caused by short-term weather disturbances (such as sudden rainfall and gusts) and random electricity consumption behavior of users (such as the start-up of temporary high-power equipment), corresponding to "component 2" in the figure; Medium-frequency intraday cycle component: energy accounts for 20%-30%, mainly reflecting the intraday peak and valley electricity consumption patterns (such as the industrial morning peak of 8:00-10:00 and the commercial evening peak of 18:00-20:00), corresponding to "component 1" in the figure; Low-frequency seasonal trend component: accounting for 55%-70% of energy, mainly reflecting the long-term load trend caused by seasonal meteorological changes (summer cooling, winter heating), corresponding to "component 3" in the figure.

[0043] Step 3: PSO Seasonal Differentiation Hyperparameter Optimization This paper implements PSO optimization using the Python programming language combined with the scikit-learn and TensorFlow frameworks. The optimization objective is the validation set determination coefficient (R²). 2 To maximize this, follow these steps: 3.1 PSO Algorithm Parameter Settings Core parameters: number of particles = 20, maximum number of iterations = 40, inertia weight = 0.9, cognitive coefficient c1 = 2.0, social coefficient c2 = 2.0; Search space constraints: Particle positions (hyperparameters to be optimized) must be within a preset range; if they exceed the range, they will be automatically clipped to the boundary value.

[0044] 3.2 Hyperparameters to be optimized and search range The hyperparameters to be optimized and the search range are shown in Table 1.

[0045] Table 1. Hyperparameters to be optimized and search range 3.3 Seasonal Differentiation Optimization Results The optimal hyperparameter combinations for each season were obtained through PSO iterative optimization, as shown in Table 2: Table 2. Schematic diagram of optimal hyperparameter combinations for each season The results show that the PSO algorithm effectively achieves seasonal differentiation optimization: In summer, due to the severe load fluctuations, LSTM adopts a lower learning rate (0.0001) to avoid training oscillations, and RF adopts a larger maximum depth (20) to mine complex feature associations; In spring / autumn, the load is stable, the number of LSTM stacking layers is increased to 4 to capture long temporal dependencies, and the RF depth is fixed at 17 to balance accuracy and efficiency.

[0046] Step 4: Sub-model prediction Based on the seasonally optimal hyperparameters obtained in step 3, an LSTM temporal dependency sub-model and a RF multi-feature association sub-model were constructed, respectively. Specifically, the Python TensorFlow framework (corresponding to LSTM) and the scikit-learn framework (corresponding to RF) were used for implementation. The prediction results are as follows: Figures 7a-10c As shown, the prediction results for spring component 1 are as follows: Figure 7a As shown, the prediction results for spring component 2 are as follows: Figure 7b As shown, the prediction results for spring component 3 are as follows: Figure 7c As shown, the prediction results for summer component 1 are as follows: Figure 8a As shown, the prediction results for summer component 2 are as follows: Figure 8b As shown, the prediction results for summer component 3 are as follows: Figure 8c As shown, the prediction results for component 1 in autumn are as follows: Figure 9a As shown, the prediction results for autumn component 2 are as follows: Figure 9b As shown, the prediction results for autumn component 3 are as follows: Figure 9c As shown, the prediction results for winter component 1 are as follows: Figure 10a As shown, the prediction results for winter component 2 are as follows: Figure 10b As shown, the prediction results for winter component 3 are as follows: Figure 10c As shown, the specific prediction operation is as follows: 4.1 LSTM Temporal Dependency Sub-model (Adapted to Low-Frequency Seasonal Trend Components) Input data: historical data of low-frequency seasonal trend components (the first 24 time steps, i.e., 12 hours) + meteorological / time / electricity price characteristics of the corresponding time period; Model structure: Input layer → LSTM layer (the number of neurons and stacked layers are set according to seasonal hyperparameters) → Fully connected layer (1 output node); Training configuration: Adam optimizer, mean squared error (MSE) loss function, batch size = 32, training epochs = 50, early stopping strategy enabled (training stops if the validation set loss does not decrease for 5 consecutive epochs); Prediction Output: Predict the low-frequency components of the test set and output the predicted value sequence of the low-frequency components at half-hour intervals. This corresponds to the "low-frequency component - actual value vs. predicted value" curve in the figure. Fit R. 2 =0.968 (summer) and 0.978 (spring), demonstrating LSTM's ability to accurately capture long-term trends.

[0047] 4.2 RF Multi-Feature Correlation Sub-model (Adapted to High / Medium Frequency Components) Input data: historical data of high / medium frequency components (the first 12 time steps, i.e., 6 hours) + complete multi-dimensional input feature set; Model training: Bootstrap resampling is used to generate 100 sample subsets, and one decision tree is trained for each subset; when splitting the decision tree, √(total number of features) candidate features are randomly selected, based on... coefficient (calculation formula is) Where D is the node sample set, Determine the optimal split point (based on the probability that a sample belongs to the k-th class); Prediction Output: Predict the high-frequency and mid-frequency components of the test set separately, outputting a sequence of predicted values ​​for both categories (at half-hour intervals). The prediction result is the arithmetic mean of all decision tree predictions, corresponding to the curves "High-frequency component - actual value vs. predicted value" and "Mid-frequency component - actual value vs. predicted value" in the figure. Rs is fitted. 2 The values ​​were 0.952 (high frequency) and 0.958 (medium frequency), respectively, to verify the effectiveness of RF in mining multi-feature associations.

[0048] Step 5: Weighted Integration Weight coefficients are calculated based on the mean absolute percentage error (MAPE) of each sub-model on the test set. The prediction results of the three components are then weighted and integrated. The specific operation is as follows: Weight coefficient calculation: The weight coefficient is inversely proportional to the MAPE of the sub-model. Let the MAPE of the sub-model corresponding to the three components be respectively... (Low frequency) (High frequency) (Medium frequency), then the weighting coefficient , , Similarly, taking the summer test set as an example, =0.93%, =1.20% =1.00%, calculated as follows =0.42、 =0.28、 =0.30; Final forecast calculation: According to the formula "Final short-term load forecast = ×Low-frequency prediction value ×High-frequency prediction value The calculation of the "×medium frequency prediction value" outputs the final prediction result for the half-hour interval, as shown below. Figures 11a-11d As shown, the load forecast fitting pair for spring is as follows: Figure 11a As shown, the summer load forecast fitting example is... Figure 11b As shown, the load forecast fitting comparison for autumn is as follows: Figure 11c As shown, the load forecast fitting comparison for winter is as follows: Figure 11d As shown, the final predicted R 2 (All ≥ 0.969).

[0049] III. Implementation Results Verification 1. Prediction accuracy metrics (test set) Table 3. Comparison of Predictive Performance between This Method and Traditional Models Using the coefficient of determination (R) 2 The mean absolute percentage error (MAPE) and root mean square error (RMSE) were used as evaluation metrics to compare the predictive performance of this method with traditional models (PSO, LSTM, RF). The results are shown in Table 3.

[0050] 2. Error Distribution Analysis Figure 11 shows a comparison of the prediction error distribution of this method, presented in box plot form. The vertical axis represents MAPE (unit: %), and the horizontal axis represents the model type, including five types of models: VMD-PSO-LSTM-RF, PSO-LSTM-RF, LSTM-RF, LSTM, and RF. Figure 12 In the table, orange corresponds to the error distribution region of LSTM, green to RF, blue to LSTM-RF, purple to PSO-LSTM-RF, and red to VMD-PSO-LSTM-RF. Key information on the error distribution is shown in Table 4. Table 4. Schematic diagram of key information on error distribution Error characteristics of this method: median MAPE is about 0.8%, box height (Q3-Q1) is only 0.5%, there are no outliers, and the error is concentrated and stable; Single model error characteristics: The median MAPE of a single LSTM is 2.25%, the box height is 1.2%, and there are multiple high error outliers (MAPE > 4.5%); the median MAPE of a single RF is 2.1%, the box height is 1.0%, there are fewer outliers but the overall error is relatively high.

[0051] 2.2 Error Distribution Conclusion Depend on Figure 12 As shown in the table above, the prediction error of this method is significantly better than that of the traditional model: The seasonal fluctuation of MAPE is only 0.36% (0.57% in winter and 0.93% in summer), which is much lower than that of a single model; The error distribution concentration is improved by more than 60% (0.5% for box height vs. 1.0%-1.2% for single model), the model is more resistant to random disturbances, and it is more suitable for the electricity spot market's demand for forecast stability.

[0052] 3. Summary of Implementation Results Accuracy advantage: This method has cross-seasonal R2 All values ​​remained above 0.969, with MAPE below 1%. Compared to single LSTM and RF models, R... 2 The improvement is 5%-13%, and the MAPE is reduced by more than 50%, which fully verifies the technical effectiveness of "VMD decomposition + PSO optimization + multi-model collaboration". Robustness advantages: Combining Figure 12 The error distribution shown is similar to that shown below. Figure 11a The seasonal fitting results shown by -d indicate that the MAPE fluctuation is only 0.36% in each season, and the predicted curve closely matches the actual load curve. For example, the fitted R value for the extreme high temperature week in summer... 2 =0.968, seasonal differentiation strategy effectively improves cross-seasonal stability; Practical advantages: Half-hour interval forecast results can be directly connected to the "daily trading-real-time dispatch" demand of the electricity spot market. The generation side can formulate unit start-up and shutdown plans based on the forecast, and the user side can use peak and valley forecasts to participate in demand response and reduce electricity costs.

[0053] IV. Key Precautions Data alignment: All types of data (load, weather, electricity price) must be strictly aligned according to the timestamp to avoid distortion of feature association due to time deviation, and to ensure that "load data at 14:00" corresponds one-to-one with "temperature and electricity price data at 14:00"; Hyperparameter boundary control: When optimizing PSO, hyperparameters need to be limited to a reasonable range (e.g., the LSTM learning rate should not be lower than 0.0001) to prevent the model from diverging during training. When the learning rate is lower than 0.0001 in summer, the model is prone to oscillation at the peak load. Early stopping strategy enabled: LSTM training requires enabling an early stopping strategy. The training stops if the validation set loss does not decrease for five consecutive epochs to avoid overfitting. This strategy can improve the validation set R0. 2 Increase by 2%-3%; Weight coefficient update: When changing the application scenario (such as users in different regions or industries), the sub-model MAPE needs to be recalculated and the weight coefficients updated. For example, industrial users have a higher proportion of high-frequency component errors, so the weight of the RF sub-model needs to be appropriately increased. Example 2 This embodiment employs a short-term load forecasting system for the electricity spot market across seasons, characterized in that the system includes: The data acquisition and processing module is configured to acquire raw power load data and meteorological data and time feature data for the corresponding time period, construct a multi-dimensional input feature set, and construct a training set and a validation set based on the multi-dimensional input feature set; Using raw electricity load data at half-hour intervals for users in the electricity spot market, combined with meteorological data and time feature data for the corresponding time period, a multi-dimensional input feature set is constructed after processing missing data values ​​using the collaborative linear interpolation method. The data acquisition and processing module includes a data receiving unit, a data preprocessing unit, and a feature fusion unit; The data receiving unit is used to import raw power load data, meteorological data, and time characteristic data and convert them into a unified format; The data preprocessing unit is used to perform cooperative linear interpolation to process missing values ​​and remove outlier data; The feature fusion unit is used to align the preprocessed data according to the time dimension and splice them together to form a multi-dimensional input feature set.

[0054] The mode decomposition and feature allocation module is configured to use the variational mode decomposition algorithm to adaptively decompose the original power load data in the multi-dimensional input feature set to obtain three types of mode components: high-frequency random fluctuation component, medium-frequency intraday cycle component, and low-frequency seasonal trend component. Based on the correspondence of the three types of mode components, the multi-dimensional input feature set is divided into random fluctuation input feature set, intraday cycle input feature set, and seasonal trend input feature set. The modal decomposition and feature allocation module includes a parameter configuration unit, a decomposition execution unit, and a component output unit; The parameter configuration unit is used to preset or adjust parameters such as the number of decomposition modes and penalty factor in VMD; The decomposition execution unit is used to run the VMD decomposition logic to achieve multi-scale separation of the original load sequence; The component output unit is used to store and output three types of modal components, providing input for subsequent sub-model prediction.

[0055] The model optimization and construction module is configured to use the particle swarm optimization algorithm to maximize the coefficient of determination on the validation set. Based on the training set, the hyperparameters of the Long Short-Term Memory Network and the Random Forest are optimized seasonally. The optimized Long Short-Term Memory Network is used to construct a temporal dependency prediction sub-model, and the optimized Random Forest is used to construct a multi-feature association prediction sub-model. The PSO algorithm is used to determine the validation set coefficient of determination (R²). 2 With the goal of maximizing, seasonally differentiated optimization is performed on the hyperparameters of LSTM and RF; The parameter optimization module includes a parameter definition unit, a PSO execution unit, and a seasonal scene recognition unit; The parameter definition unit is used to determine the hyperparameters to be optimized and the search range for LSTM and RF; The PSO execution unit is used to run the particle swarm optimization logic and iteratively search for the optimal combination of hyperparameters. The seasonal scene recognition unit is used to adjust and optimize the target weights based on the different seasonal load characteristics and output seasonal-specific hyperparameters.

[0056] The seasonal scene recognition unit is configured as follows: Extract month information for the date to be predicted from time feature data; Based on the monthly information, the current period to be predicted is determined to be a summer scenario, a winter scenario, or a spring / autumn scenario; Based on the determined scenario type, the corresponding long short-term memory network hyperparameter search interval and random forest hyperparameter search interval are retrieved from the system database, and the search interval is passed as a constraint condition to the particle swarm optimization algorithm.

[0057] The multi-component prediction module is configured to input the seasonal trend input feature set into the time-series dependent prediction sub-model and output the first prediction result, and input the random fluctuation input feature set and the intraday cycle input feature set into the multi-feature association prediction sub-model and output the second prediction result. To address the feature differences among the three modal components, an optimized LSTM is used to construct a time-dependent prediction sub-model, and an optimized RF is used to construct a multi-feature association prediction sub-model, to obtain the prediction results for each modal component. The multi-component prediction module includes an LSTM training unit, an RF training unit, and a component prediction unit. The LSTM training unit is used to train the temporal dependency sub-model based on the optimized hyperparameters to capture the temporal dependencies of low-frequency components. The RF training unit is used to train a multi-feature association sub-model based on optimized hyperparameters to discover the association patterns between high / medium frequency components and multiple features; The component prediction unit is used to call the trained sub-model and output the prediction results of each modal component.

[0058] The results integration module is configured to weight and integrate the first and second forecast results to obtain the short-term load forecast results.

[0059] The prediction results of each modal component are weighted and integrated after normalizing the average absolute percentage error of each sub-model and determining the weighting coefficients to obtain the final short-term load forecast. The results integration module includes a precision calculation unit, an integration calculation unit, and a results output unit. The accuracy calculation unit is used to calculate the MAPE of each sub-model and determine the weighting coefficients; The integrated computing unit is used to calculate the final short-term load forecast using a weighted formula; The results output unit is used to output the final prediction results in tabular or graphical form.

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A short-term load forecasting method for the electricity spot market across seasons, characterized in that, The method steps include: S1. Collect raw power load data and corresponding meteorological and temporal feature data for the same period, construct a multi-dimensional input feature set, and build a training set and a validation set based on the feature set; S2. Using the variational mode decomposition algorithm, the original power load data in the multi-dimensional input feature set is decomposed into three types of mode components: random fluctuation component, intra-day periodic component, and seasonal trend component. Based on the correspondence of the three types of mode components, the multi-dimensional input feature set is divided into random fluctuation input feature set, intra-day periodic input feature set, and seasonal trend input feature set. S3. Using the particle swarm optimization algorithm, with the goal of maximizing the coefficient of determination on the validation set, the hyperparameters of the Long Short-Term Memory Network and the Random Forest are seasonally differentiated and optimized based on the training set. The optimized Long Short-Term Memory Network is used to construct a time-dependent prediction sub-model, and the optimized Random Forest is used to construct a multi-feature association prediction sub-model. S4. Input the seasonal trend input feature set into the time-series dependent prediction sub-model and output the first prediction result. Input the random fluctuation input feature set and the intraday cycle input feature set into the multi-feature association prediction sub-model and output the second prediction result. S5. The first and second forecast results are weighted and integrated to obtain the short-term load forecast result.

2. The short-term load forecasting method for the electricity spot market across seasons as described in claim 1, characterized in that, When constructing the multi-dimensional input feature set in S1, the process also includes handling missing data values. Specifically, a collaborative linear interpolation method is used to fill in the missing values ​​of the original power load data. The specific process of the collaborative linear interpolation method includes: The interpolation base value is calculated based on the effective load values ​​adjacent to the missing time in the original power load data. By introducing a preset electricity price correction factor, and combining the average electricity price of adjacent time periods, the electricity price of the missing time period, and the preset electricity price correction factor, the interpolation base value is corrected to obtain the corrected missing load value. The missing load values ​​are filled in to the missing moments of the original power load data to form a complete load data sequence; The 3σ criterion is used to detect the complete load data sequence. The mean μ and standard deviation σ of the data sequence are calculated, and data that exceed the interval [μ-3σ, μ+3σ] are identified as outliers and removed.

3. The short-term load forecasting method for the electricity spot market across seasons according to claim 1, characterized in that, The parameter configuration of the variational mode decomposition algorithm includes: Set the number of decomposed modes K=3, the penalty factor α=2000, the noise tolerance τ=0, and do not force the decomposition of DC components; When using the variational mode decomposition algorithm to decompose the original power load data into three types of modal components, the original power load data is adaptively separated according to the frequency from high to low or from low to high. Among them, the random fluctuation component corresponds to the mode with the highest frequency; the intra-day periodic component corresponds to the intermediate frequency mode; and the seasonal trend component corresponds to the mode with the lowest frequency.

4. The short-term load forecasting method for the electricity spot market across seasons according to claim 1, characterized in that, In S3, the hyperparameters of the Long Short-Term Memory network specifically include the number of hidden layer neurons, the number of stacked layers, and the learning rate; the hyperparameters of the Random Forest specifically include the number of decision trees and the maximum depth.

5. A short-term load forecasting method for the electricity spot market across seasons, as described in claim 4, is characterized in that... In S3, when using the particle swarm optimization algorithm for seasonal difference optimization, the specific configuration includes: The particle swarm optimization algorithm is set to have 20 particles and a maximum number of iterations of 40. When optimizing the Long Short-Term Memory network, the optimization search range for the number of hidden layer neurons is set to [32, 256], the optimization search range for the number of stacked layers is set to [1,5], and the optimization search range for the learning rate is set to [0.0001, 0.01]. When optimizing the Random Forest, the optimization search range for the number of decision trees is set to [50, 150], and the optimization search range for the maximum depth is set to [3, 25].

6. The short-term load forecasting method for the electricity spot market across seasons according to claim 1, characterized in that, The seasonal differentiation optimization in S3 satisfies the following seasonal adaptation rules: In the summer scenario, the learning rate of the Long Short-Term Memory network obtained by the particle swarm optimization algorithm is set to the interval [0.0001, 0.01), and the maximum depth of the random forest is set to the interval [19, 20]. In spring or autumn scenarios, the number of stacked layers of the long short-term memory network obtained by the particle swarm optimization algorithm is set to the interval [2, 4] and the learning rate is set to 0.

01. The maximum depth of the random forest is set to 17. In the winter scenario, the number of stacked layers of the Long Short-Term Memory network obtained by the Particle Swarm Optimization algorithm is set to the interval [3, 4], and the maximum depth of the Random Forest is set to the interval [19, 20].

7. A short-term load forecasting method for the electricity spot market across seasons as described in claim 1, characterized in that, In S4, the specific prediction logic of each sub-model is as follows: For the time-dependent prediction sub-model, the seasonal trend input feature set is used as input. The gating mechanism of the long short-term memory network is used to extract the long-term memory features and short-term evolution rules of the load trend sequence in the time dimension. Based on the long-term memory features and short-term evolution rules, the load trend at the next moment is extrapolated in time, thereby outputting the first prediction result. For the multi-feature association prediction sub-model, the random fluctuation input feature set and the intraday cycle input feature set are used as inputs. The ensemble decision tree structure of random forest is used to mine the nonlinear coupling relationship between meteorological data and time feature data and load fluctuation components. Based on the coupling relationship, the load fluctuation value at the next moment is regressed and predicted, thereby outputting the second prediction result.

8. A short-term load forecasting method for the electricity spot market across seasons as described in claim 1, characterized in that, The specific process of weighted integration of the first and second prediction results in S5 includes: Using validation set data, the mean absolute percentage error of the time-dependent prediction sub-model and the multi-feature association prediction sub-model were calculated respectively. The prediction accuracy of each sub-model is evaluated based on the mean absolute percentage error, and the weighting coefficient of the prediction result of each sub-model is determined according to the prediction accuracy of each sub-model. Based on the weighting coefficients, the first and second prediction results are weighted and summed to obtain the short-term load prediction result.

9. A short-term load forecasting system for the electricity spot market across seasons, characterized in that, The system operates using a short-term load forecasting method for the electricity spot market across seasons, as described in any one of claims 1-8. The system comprises: The data acquisition and processing module is configured to acquire raw power load data and meteorological data and time feature data for the corresponding time period, construct a multi-dimensional input feature set, and construct a training set and a validation set based on the multi-dimensional input feature set; The modal decomposition and feature allocation module is configured to use the variational modal decomposition algorithm to decompose the original power load data in the multi-dimensional input feature set into three types of modal components: random fluctuation component, intra-day periodic component, and seasonal trend component. Based on the three types of modal components, the multi-dimensional input feature set is divided into random fluctuation input feature set, intra-day periodic input feature set, and seasonal trend input feature set. The model optimization and construction module is configured to use the particle swarm optimization algorithm to maximize the coefficient of determination on the validation set. Based on the training set, the hyperparameters of the Long Short-Term Memory Network and the Random Forest are optimized seasonally. The optimized Long Short-Term Memory Network is used to construct a temporal dependency prediction sub-model, and the optimized Random Forest is used to construct a multi-feature association prediction sub-model. The multi-component prediction module is configured to input the seasonal trend input feature set into the time-series dependent prediction sub-model and output the first prediction result, and input the random fluctuation input feature set and the intraday cycle input feature set into the multi-feature association prediction sub-model and output the second prediction result. The results integration module is configured to weight and integrate the first and second forecast results to obtain the short-term load forecast results.

10. A short-term load forecasting system for the electricity spot market across seasons according to claim 9, characterized in that, The model optimization and construction module includes a pre-built seasonal scene recognition unit. The seasonal scene recognition unit is configured as follows: Extract the month information of the date to be predicted from the time feature data; Based on the month information, the current period to be predicted is determined to be a summer scene, a winter scene, or a spring / autumn scene; Based on the determined scenario type, the corresponding long short-term memory network hyperparameter search interval and random forest hyperparameter search interval are retrieved from the system database, and the search interval is passed as a constraint condition to the particle swarm optimization algorithm.

Images