Power grid load forecasting method based on adaptive VMD-RF-LSTM

Abstract

The present application belongs to the technical field of energy data processing, and aims at the problems of insufficient expression ability, robustness and time rhythm adaptability of existing power grid load prediction, and proposes a power grid load prediction method based on adaptive VMD-RF-LSTM: data set division is performed to generate exogenous characteristics; the actual power grid load after data processing is decomposed into intrinsic mode function and residual term by using adaptive variational mode decomposition; after obtaining standardized data, a single-level LSTM regression model is trained for each subsequence; for the intrinsic mode function, a single-level LSTM regression model is directly used for training to obtain a prediction result; for the residual term, a prediction result is obtained by using RF and a single-level LSTM regression model; all prediction results are added to obtain a total power grid load prediction result. The present application improves the expression ability, robustness and sensitivity to time rhythm of the prediction model through the power grid load prediction method based on adaptive VMD-RF-LSTM.

Description

Technical Field

[0001] This invention relates to the field of energy data processing technology, and specifically to a power grid load forecasting method based on adaptive VMD-RF-LSTM. Background Technology

[0002] In modern power systems, time-series data has become a core foundation for real-time monitoring, analysis, and decision-making. Data from multiple sources, including system load, transformer oil temperature, cross-sectional power flow, and renewable energy output, sequentially depict the dynamic evolution of load fluctuations, equipment thermal-electric coupling, and transmission bottlenecks. Compared to static statistics, time-series models can capture nonlinear, seasonal, and short-term fluctuations, revealing latent trends and abnormal behaviors in advance, providing crucial information for dispatch centers to achieve efficient rolling dispatch, capacity planning, fault early warning, and safety assessment.

[0003] Especially under the "dual carbon" context, the high proportion of renewable energy grid connection significantly enhances the uncertainty of both source and load. The accuracy of system load forecasting directly affects reserve capacity configuration, energy storage charging and discharging strategies, and cross-sectional power flow control, and is a key technical support for ensuring grid security margin and reducing operating costs and carbon emissions. Therefore, it is urgent to construct a unified time series framework to perform collaborative forecasting and uncertainty quantification of multi-source data, providing a reliable boundary for the stable and economical operation of the power system.

[0004] Currently, data-driven methods such as Long Short-Term Memory (LSTM), Random Forest (RF), and XGBoost have become mainstream tools for time series analysis. However, single models often struggle to fully integrate exogenous driving factors, lack robustness to extreme events, and have limited ability to capture the spatiotemporal coupling relationship between weather, renewable energy output, and load. Although some multimodal or hybrid methods (such as EMD-BiLSTM and stacked autoencoder-XGBoost) have improved prediction performance in certain scenarios, they still generally suffer from problems such as mode mixing, endpoint effects, and incomplete separation of multi-scale oscillations, failing to fully utilize the effective signals in the residual information. Existing methods still exhibit sluggish responses to high-frequency disturbances such as sudden changes in wind and solar power output and load spikes, making it difficult to provide reliable "next hour system load" prediction boundaries, thus affecting reserve capacity retention and cross-sectional power flow safety control, and failing to meet the dispatch center's need for rapid correction on a second-to-minute scale.

[0005] To address the problems of existing technical methods, such as significant mode mixing and endpoint effects, insufficient response to extreme events and high-frequency disturbances, inadequate fusion of multi-source time-series information, and lack of ability to quantify prediction uncertainty, this invention proposes a power grid load forecasting method based on adaptive VMD-RF-LSTM. The aim is to improve the accuracy and robustness of power grid load forecasting and provide reliable uncertainty boundaries for scheduling decisions. Summary of the Invention

[0006] In view of this, in order to address the shortcomings of existing prediction models in terms of expressive power, robustness, and adaptability to temporal rhythms, this invention provides a power grid load forecasting method based on adaptive VMD-RF-LSTM. This method aims to simultaneously improve prediction accuracy and model robustness. By adaptively decomposing and denoising the load sequence, explicitly modeling extreme events, and quantifying the uncertainty of the prediction results, it effectively enhances the ability to capture short-term fluctuations. This provides a reliable basis for the dispatch center to adjust the output curve on a second-to-minute scale, optimize energy storage charging and discharging strategies, and correct reserve configurations, thereby enhancing the stability margin of the power grid from the source of system operation.

[0007] This invention provides a power grid load forecasting method based on adaptive VMD-RF-LSTM, comprising: Step 110: Input multi-source time series data of the power system and perform preprocessing; the multi-source time series data includes at least actual generation data, dispatch plan boundary data, actual grid load and day-ahead electricity price; divide the preprocessed multi-source time series data into datasets based on the same time series according to a preset ratio: training set, validation set and test set; Step 120: Construct features for the multi-source time series data after the dataset is divided, and generate exogenous features for the training set, validation set, or test set respectively; the exogenous features include at least multi-dimensional time features and dataset-inherent parameter features; concatenate the multi-source time series data other than the actual power grid load with the corresponding exogenous features column by column to form an input data matrix; Step 130: Using adaptive variational mode decomposition, the preprocessed actual power grid load is adaptively decomposed into... Subsequences, including Each intrinsic mode function and one residual term; each subsequence contains a dataset partition with the same time sequence as the actual grid load before decomposition; Step 140: Perform standardization and nonlinear transformation on each column of the input data matrix and each subsequence to obtain the corresponding standardized data; Step 150: Construct and train a single-level long short-term memory network regression model for each subsequence based on the standardized data of the training set; if the subsequence is a residual term, filter processing is performed first to suppress noise interference; use the corresponding trained single-level long short-term memory network to predict the intrinsic mode function or the filtered residual term, and then obtain the prediction result for each subsequence after inverse standardization and inverse linear transformation. Step 160, will The prediction results of each subsequence are summed at the same time nodes of the time series to obtain the final power grid load prediction result.

[0008] Preferably, the power grid load forecasting method based on adaptive VMD-RF-LSTM further includes an iterative optimization step based on external variable prediction feedback: Step 1: Calculate the correlation index between external variables and corresponding variables of actual power grid load in the original multi-source time series data; select external variables whose correlation index is greater than a preset correlation threshold as strongly correlated external variables; the external variable refers to the corresponding variable of any other multi-source time series data besides the actual power grid load; the coherence index is the Pearson correlation coefficient or mutual information value. Step 2: For each strongly correlated external variable, perform mode decomposition according to the adaptive variational mode decomposition method in step 130 to obtain the corresponding intrinsic mode function and residual term; then, according to the process of steps 140 to 160, obtain the predicted value of the strongly correlated external variable in the future preset time period. Step 3: Using the predicted values ​​of the strongly correlated external variables, update the time series data of the corresponding strongly correlated external variables in the original multi-source time series dataset during the preset time period to obtain the updated multi-source time series dataset. Step 4: Based on the updated multi-source time series dataset, return to step 120, and execute steps 120 to 160 sequentially to obtain the updated power grid load forecast results; Step 5: Repeat steps 2 to 4 until the preset iteration termination condition is met, thereby realizing the rolling optimization prediction of power grid load.

[0009] In summary, this invention provides a power grid load forecasting method based on adaptive VMD-RF-LSTM. Compared with existing technologies, the method and apparatus of this invention have the following advantages: (1) By using adaptive VMD to decompose complex non-stationary load sequences into stationary subsequences, the modeling difficulty of each sub-model is reduced; at the same time, the LSTM model fully integrates the multi-source time series features of the system, enhancing the understanding of the driving factors of load changes.

[0010] (2) Using Median-MAD normalization and inverse hyperbolic sine transform to process data can effectively suppress outliers and noise interference, and improve the stability of the model in extreme weather or market fluctuation scenarios.

[0011] (3) All operations that learn parameters from the data (such as VMD decomposition and standardization) are strictly limited to the training set, and the learned rules are consistently applied to the validation set and the test set, which fundamentally avoids data leakage and ensures the authenticity of the evaluation results and the generalization ability of the model.

[0012] (4) By using the counterintuitive operation of “updating the input with the prediction results”, the barrier of “one-way data flow” in the traditional prediction model is broken, and a closed loop of “data-prediction-data” is constructed, which enables the prediction system to have the ability to self-correct and perceive the environment. The final output of the power grid load prediction result is not only a mapping of historical data, but also an adaptive response to changes in key variables in the future.

[0013] Further experimental verification shows that the power grid load forecasting method provided by this invention can provide the dispatch center with a more reliable "next hour" or "day-ahead" load forecast boundary, and directly support advanced applications such as reserve capacity configuration, energy storage dispatch and cross-sectional power flow safety verification. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a power grid load forecasting method based on adaptive VMD-RF-LSTM in one embodiment of the present invention. Figure 2 This is a schematic diagram of a power grid load forecasting algorithm framework based on adaptive VMD-RF-LSTM in one embodiment of the present invention; Figure 3 This is a schematic diagram of a single-stage LSTM regression model algorithm architecture in one embodiment of the present invention; Figure 4 This is a schematic diagram of a power grid load forecasting algorithm framework based on adaptive VMD-RF-LSTM that adds a feedback process by updating day-ahead electricity price data in one embodiment of the present invention. Figure 5 This diagram illustrates the comparison of RMSE (Reference to Model Evaluation Index) results obtained by training different models based on a 4-year power system dataset from country F in the experiments of this invention. The horizontal axis represents the baseline models selected for comparison: RF, XGBoost, LSTM, Multi-CNN, CNN-LSTM, Encoder-Decoder, and VMD-LSTM. The last bar chart represents VMD-RF-LSTM used in this invention. The vertical axis represents the RMSE (Reference to Model Evaluation Index). Figure 6 This diagram illustrates the comparison of MAPE (Model Evaluation Index) results obtained by training different models using a four-year power system dataset from country F in the experiments of this invention. The horizontal axis represents the baseline models selected for comparison: RF, XGBoost, LSTM, Multi-CNN, CNN-LSTM, Encoder-Decoder, and VMD-LSTM. The last bar represents VMD-RF-LSTM used in this invention. The vertical axis represents the MAPE index, expressed as a percentage (%). Figure 7This diagram illustrates the comparison of SMAPE (Signature Mapping Efficiency) evaluation results obtained using different models trained on a 4-year power system dataset from country F in this invention. The horizontal axis represents the baseline models selected for comparison: RF, XGBoost, LSTM, Multi-CNN, CNN-LSTM, Encoder-Decoder, and VMD-LSTM. The last bar represents VMD-RF-LSTM used in this invention. The vertical axis represents the model evaluation metric SMAPE, expressed in %. Figure 8 This diagram illustrates the comparison of results obtained by training different models using a 4-year power system dataset from country F in the experiments of this invention, with an evaluation index of R². The horizontal axis represents the baseline models selected for comparison: RF, XGBoost, LSTM, Multi-CNN, CNN-LSTM, Encoder-Decoder, and VMD-LSTM. The last bar represents VMD-RF-LSTM used in this invention. The vertical axis represents the model evaluation index R². Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0016] To address the shortcomings of existing technologies, the purpose of this invention is to more accurately capture the fluctuation patterns and trading activity of price series at different times, thereby improving the model's expressive power, robustness, and sensitivity to temporal rhythms. Specifically, it systematically improves time series processing and exogenous information integration, forming a time-centric multimodal prediction framework: a power grid load prediction model based on adaptive VMD-RF-LSTM (VMD with Residual-Filtered LSTM, a long short-term memory network based on adaptive variational mode decomposition and residual term filtering). This multimodal prediction framework first uses adaptive variational mode decomposition (VMD) to adaptively decompose the complex nonlinear, non-stationary actual power grid load time series into several intrinsic mode function (IMF) components with finite bandwidth and a residual term, serving as subsequences of the actual power grid load time series. The IMF components have a center frequency characteristic in the frequency domain, with each IMF representing an oscillation mode within a specific frequency band in the original signal (data). This process is similar in effect to decomposing a complex signal into a set of frequency-variable "harmonic components." Then, prediction is performed at the subsequence level, and the residual terms obtained from the decomposition are filtered and predicted again. Finally, the prediction results of each subsequence are summed and integrated to obtain the final comprehensive prediction value. This model architecture based on "decomposition-prediction-integration" uses an adaptive decomposition algorithm to decouple different modes in complex signals, and tailors the most suitable prediction strategy for the uniqueness of each mode, thereby achieving accurate "divide and conquer" and finally integrating into superior overall prediction performance. The entire process focuses on the multi-scale characteristics and robustness of actual power grid load time series, thereby improving the accuracy and stability of prediction.

[0017] In one embodiment, such as Figure 1 As shown, this invention provides a power grid load forecasting method based on adaptive VMD-RF-LSTM, comprising: Step 110: Input multi-source time series data of the power system and perform preprocessing; the multi-source time series data includes at least actual generation data, dispatch plan boundary data, actual grid load and day-ahead electricity price; divide the preprocessed multi-source time series data into datasets based on the same time series according to a preset ratio: training set, validation set and test set; Step 120: Construct features for the multi-source time series data after the dataset is divided, and generate exogenous features for the training set, validation set, or test set respectively; the exogenous features include at least multi-dimensional time features and dataset-inherent parameter features; concatenate the multi-source time series data other than the actual power grid load with the corresponding exogenous features column by column to form an input data matrix; Step 130: Using adaptive variational mode decomposition, the preprocessed actual power grid load is adaptively decomposed into... Subsequences, including Each intrinsic mode function and one residual term; each subsequence contains a dataset partition with the same time sequence as the actual grid load before decomposition; Step 140: Perform standardization and nonlinear transformation on each column of the input data matrix and each subsequence to obtain the corresponding standardized data; Step 150: Construct and train a single-level long short-term memory network regression model for each subsequence based on the standardized data of the training set; if the subsequence is a residual term, filter processing is performed first to suppress noise interference; use the corresponding trained single-level long short-term memory network to predict the intrinsic mode function or the filtered residual term, and then obtain the prediction result for each subsequence after inverse standardization and inverse linear transformation. Step 160, will The prediction results of each subsequence are summed at the same time nodes of the time series to obtain the final power grid load prediction result.

[0018] Specifically, in step 110, the actual data on the power generation side includes at least: biomass power generation, fossil lignite / lignite power generation, fossil gas power generation, fossil hard coal power generation, pumped storage power consumption, runoff and small reservoir hydropower generation, reservoir hydropower generation, nuclear power generation, conventional energy power generation not otherwise listed, solar power generation, waste power generation, onshore wind power, and renewable energy power generation not otherwise listed.

[0019] The boundary data of the scheduling plan includes at least: day-ahead solar power forecast, day-ahead onshore wind power forecast, and total load forecast.

[0020] The data preprocessing includes at least missing value imputation and duplicate value handling. Figure 2 A schematic diagram of the algorithmic framework of the power grid load forecasting method based on adaptive VMD-RF-LSTM in this embodiment is given.

[0021] In real-world scenarios, large-scale time-series datasets are often collected and integrated to cover sufficiently long time spans. However, with the accumulation of factors such as the diversity of data sources, transmission interruptions, sensor failures, daylight saving time adjustments, and sampling gaps, missing values ​​inevitably appear in continuous multi-source time-series data. If these temporal missing values ​​are not handled properly, they will disrupt the continuity of the time series, destroy the temporal dependencies of features, and thus affect the training stability, prediction accuracy, and evaluation reliability of the model. Therefore, reasonable and robust strategies for missing value imputation and duplicate value handling are needed to maintain the continuity and statistical properties of the time series as much as possible.

[0022] Specifically, for any multi-source time series data ; ; This is the total number of time points contained in the time series. If a certain time point... If a missing value exists, take the previous time point. With the next time node average As a time node Data filling at the location: ; If a certain time point exists Multiple data points exist at the same location: , If the number of duplicate data points is a given, then the average of these multiple prices is taken as the price data for that time point. .

[0023] For example, if the time node is to obtain one data point per hour (hourly data), then for missing hour values, the average of the previous hour (time node) and the next hour (time node) is used to fill the gap. For multiple data points appearing repeatedly at the same time node (one hour) (such as data anomalies), the average of the multiple prices is used as the data for that time node.

[0024] Furthermore, to ensure that training, validation, and testing are strictly divided in chronological order and to avoid time pollution and information leakage, the preprocessed multi-source time series data are divided into datasets based on the same time sequence according to a preset ratio: training set, validation set, and test set.

[0025] First, the preprocessed multi-source time series data are sorted according to the time window to form a time series dataset.

[0026] Then the dataset is divided according to a preset ratio: Training set: ; Validation set: ; Test set: .

[0027] The above dataset partitioning must ensure that any data feature and label from the same day are in the same dataset partition to avoid information leakage across partitions.

[0028] In step 120, feature construction is performed on the multi-source time series data after the dataset is divided, and exogenous features are generated for the training set, validation set, or test set respectively; the exogenous features include at least multi-dimensional time features and dataset-inherent parameter features; the multi-source time series data other than the actual power grid load is concatenated with the corresponding exogenous features column by column to form an input data matrix.

[0029] Based on the timestamps themselves and the feature information contained in the original multi-source time series data, exogenous features for each dataset of each multi-source time series data are generated by designing and updating data features.

[0030] The multi-dimensional time features include at least: "hour", "weekday", "month", and "business hour" based on the work time zone. Specifically: The hour, day of the week, and month of each record are extracted using DatetimeIndex and written into new columns hour / weekday / month, respectively, providing intuitive time features for subsequent modeling. DatetimeIndex is a dedicated data structure in the Python data analysis library Pandas used for storing and manipulating timestamp sequences.

[0031] Based on the division into business hours (morning / lunch break / other days) and weekday (0 / 1 / 2 codes for Saturday, Sunday, and weekdays), two columns are generated to further depict the changes in power grid load with the work and rest patterns.

[0032] In addition, it can also include features that come with the dataset itself, such as Generationforecast and System load forecast of actual power grid load data.

[0033] Furthermore, in step 130, adaptive variational mode decomposition is used to adaptively decompose the preprocessed actual grid load into... Subsequences, including Each intrinsic mode function and one residual term; each subsequence contains a dataset partition with the same time sequence as the actual grid load before decomposition.

[0034] The Adaptive Variational Mode Decomposition (VMD) is an adaptive signal decomposition method used to decompose complex signals into several intrinsic mode function components (IMFs, hereinafter referred to as modes) with finite bandwidths. Each component is as narrow-bandwidth as possible in the frequency domain. The core idea of ​​adaptive VMD is to decompose the original data signal into several sub-signals (sub-sequences) by solving a constrained optimization problem, such that each sub-signal has a concentrated bandwidth in the frequency domain and maximizes mutual orthogonality. The parameters required for training in this process include at least: : Bandwidth constraint parameters (bandwidth weight / smoothness parameters) for each mode; : The number of intrinsic mode functions obtained by decomposition.

[0035] In addition, a convergence adjustment parameter can be preset. .

[0036] First, adaptive variational mode decomposition is used to adaptively decompose the preprocessed actual power grid load into multiple ( (number) subsequences, including There are one intrinsic mode function and one residual term. Specifically, the formula (mathematical form of the optimization problem) for the above decomposition process is described as follows: For the pre-processed actual power grid load The goal is to break it down into Individual eigenmode functions And the residual term, such that: , ; in, It is the first Individual eigenmode functions The corresponding center frequency, , For unit impact function, This represents the convolution operation. It is the imaginary unit. It calculates the gradient with respect to time. Represents the square of the L2 norm; This indicates that the intrinsic mode functions The spectrum from The frequency is near the center frequency, and the spectrum is shifted to the zero frequency center, converting the bandpass signal into a low-pass signal; The Hilbert transform is performed on the intrinsic mode functions (IMFs) after the transfer center frequency using convolution operations to obtain the analytic signals of the IMFs. These analytic signals contain only positive frequency components, facilitating the analysis of frequency characteristics; constraints... Ensure that all subsequences (IMFs and residuals) are perfectly reconstructed into the original signal. This ensures the completeness of the decomposition and prevents information loss.

[0037] The constrained optimization problem described above is typically transformed into an unconstrained problem by introducing a quadratic penalty factor and the Lagrange multiplier method. Then, iterative solutions are performed using the alternating direction multiplier method, where each eigenmode function is updated alternately in each iteration. Its center frequency And Lagrange multipliers, until convergence. A major challenge of standard VMD is that the parameters need to be pre-defined, especially the number of intrinsic mode functions. and bandwidth constraint parameters Improper settings can lead to mode aliasing (mixing of different intrinsic mode function components) or over-decomposition / under-decomposition.

[0038] By solving the above optimization problem, VMD can handle complex real-world power grid loads. (The original signal) is adaptively decomposed into The most compact eigenmode function with the smallest bandwidth This ensures that each IMF carries oscillation information at a specific scale from the original signal, and that the frequency overlap between IMFs is minimized. Subsequent optimization of parameters will be achieved through a search algorithm. and This decomposition achieves the global optimum under the current signal, providing the highest quality and "cleanest" subsequence components for subsequent feature extraction and prediction.

[0039] Furthermore, in order to automatically find the optimal parameters (including at least the bandwidth constraint parameters of the decomposition algorithm) and the number of intrinsic mode functions K ), through parameter space Global optimization is performed using the Sparrow Search Algorithm (SSA) to minimize quantization metrics related to signal structure, where envelope entropy is used as one of the metrics to measure the "disorder" or uncertainty of the decomposition results. Optimization can yield a cleaner combination of parameters for subsequences to support subsequent prediction, noise reduction, or feature engineering.

[0040] The process of calculating the envelope entropy includes: For each IMF obtained from VMD decomposition, the envelope signal is calculated by obtaining the instantaneous amplitude through Hilbert transform. Then, the envelope signal is standardized into a probability distribution sequence: , ; Calculate the envelope entropy for each IMF: ; Envelope entropy measures the sparsity and uncertainty of a signal's envelope. If an IMF is very pure and smooth, its envelope value varies little, resulting in a low entropy. Conversely, if an IMF contains a large amount of noise or a mixture of different modes, its envelope will fluctuate dramatically, and the normalized probability distribution will be more uniform, leading to a higher entropy. Therefore, minimizing envelope entropy (the objective is usually set as a weighted average or sum of the envelope entropies of all IMFs) means that SSA seeks the parameter combination that results in the purest and most ordered subsequences at the decomposition precipitates.

[0041] The SSA algorithm simulates the foraging and anti-predation behavior of sparrows. Through role updates of discoverer, follower, and vigilant, it efficiently explores and develops within the parameter space, ultimately finding the globally optimal or near-optimal algorithm that minimizes the envelope entropy. .

[0042] Re-analyzing the actual power grid load time series Performing a VMD decomposition yields the following results: Individual eigenmode functions and a residual term (RES): .

[0043] Then, based on the data set partitioning method of the same time sequence of the actual power grid load before decomposition, the dataset of each subsequence is divided into: training set, validation set and test set.

[0044] Specifically, in step 140, each column of the input data matrix and each subsequence are standardized and nonlinearly transformed to obtain the corresponding standardized data.

[0045] In one embodiment, the standardization is performed using Median-Mad standardization, and the nonlinear transformation is performed using an inverse hyperbolic sine transform.

[0046] Specifically, Median-Mad standardization and inverse hyperbolic sine transform are performed on each column of the input data matrix or each subsequence dataset to obtain standardized data for the corresponding training, validation, or test sets, including: (1) Using Median-Mad normalization, the time series of each column of the input data matrix or each subsequence of the entire training set is normalized. Calculate the median of the whole. Compared with absolute median MAD is a robust statistic that is better suited to outliers in a dataset than standard deviation. Compared to using the squared standard deviation of the distance from the mean, MAD has a larger weighting for larger deviations, meaning the impact of outliers cannot be ignored. With MAD, a small number of outliers will not affect the experimental results; therefore, the following standardization formula should be applied to each point in the training, validation, or test set: ; It is worth noting that, in order to avoid disclosing the data information of the validation set and the test set, the... and It was obtained by computation on the training set.

[0047] (2) Using the inverse hyperbolic sine transform (asinh), the standardized values ​​are... Perform asinh transformation: ; This transformation helps to suppress the impact of grid load spikes, making the model learn more robustly.

[0048] It is important to note that in subsequent steps, for each time series (each column of the input data matrix or each subsequence), the standardized data is input into the corresponding single-level LSTM regression model for model training and prediction, and the prediction results are output. Figure 3 As shown, it is also necessary to design inverse transformation and inverse standardization processes for the aforementioned Median-Mad standardization and asinh transform, so that the predicted values ​​output by the model during the training phase are... The original scale of the data is restored through the inverse transformation and inverse normalization, specifically: First, we transform the expression using the inverse function sinh of asinh, and obtain: ; Then perform inverse standardization to obtain the prediction results: .

[0049] It is noted that Median-Mad standardization provides robust estimates of central tendency and dispersion, effectively reducing the interference of outliers. Meanwhile, the Asinh transform preserves the linear characteristics of small values ​​while performing logarithmic compression on large values, improving numerical stability and suppressing the influence of extreme values. The combination of these two methods maintains the relative relationships of the data while enhancing the model's robustness and generalization ability under noise and drift conditions.

[0050] Further, in step 150, a single-level long short-term memory network regression model is constructed and trained for each subsequence based on the standardized data of the training set; if the subsequence is a residual term, filtering is first performed to suppress noise interference; the intrinsic mode function or the filtered residual term is predicted using the corresponding trained single-level long short-term memory network, and then the prediction result for each subsequence is obtained after inverse standardization and inverse linear transformation, including: (1) If the subsequence is an intrinsic mode function, then the standardized data is input into the single-level LSTM regression model corresponding to the subsequence, the model is trained using the standardized data of the training set, and the standardized data of the validation set / test set is input into the corresponding trained single-level LSTM regression model for prediction, so as to obtain the prediction result of the intrinsic mode function.

[0051] Specifically, for each subsequence, such as Figure 3As shown, for the input standardized data, the 100 LSTM units in the single-level LSTM regression model corresponding to the subsequence scan the entire input time series with return_sequences=True, outputting a 100-dimensional hidden state at each time step in real time. This captures both local details and long-range dependencies at once, meaning the LSTM layer returns the hidden state at each time step, with an output shape of... ,in, It refers to the time steps. `return_sequences` is a boolean parameter representing the returned sequence, used to control whether the output returned by the LSTM layer is the hidden state of the last time step. It is the batch size, which represents the number of samples input into the network at one time during model training.

[0052] Subsequently, the Flatten layer will Tensor flattening The global context vector compresses "time × feature" into a single long feature.

[0053] The long features are input into the first fully connected layer, Dense(200, ReLU), to learn high-level feature combinations in the spatial dimension; the Dense fully connected layer has 200 neurons and uses the ReLU activation function to introduce non-linearity into the neural network. The first fully connected layer takes the output of the Flatten layer... The 1D vector is mapped to a new 200-dimensional vector through a weight matrix and ReLU transformation. The role of this layer can be understood as integrating the temporal features learned from the LSTM and discovering the complex relationships between these features.

[0054] Next, a regularization layer Dropout (0.1) is used to add 10% Dropout (random deactivation, which means that in each iteration of the training process, a portion of the neurons in the network are randomly "turned off" or "dropped", and the dropout rate is 10% here) to 200 ReLU neurons to complete the high-dimensional nonlinear mapping.

[0055] The output layer Dense(1, linear) then outputs a single scalar prediction to generate the predicted value. Here, linear is a linear activation function, which means that the neuron directly outputs its weighted sum without performing a nonlinear transformation.

[0056] Finally, by performing inverse transform and inverse normalization on the predicted values ​​output by the output layer, we can obtain the predicted result of the intrinsic mode function corresponding to a subsequence, denoted as . .

[0057] (2) If the subsequence is a residual term, then the prediction results of the residual term are obtained by using variance evaluation, residual term filtering (RF) and the single-stage LSTM regression model corresponding to the residual term, including: Calculate the variance (VAR) of the residual term; If the variance VAR is not greater than a given threshold This indicates that the residuals are relatively smooth, reflecting the long-term trend. The standardized data of the residuals can be directly input into the corresponding single-level LSTM regression model for training and prediction to obtain the predicted results of the residuals. ; If the variance VAR is greater than a given threshold This indicates the presence of significant noise. Therefore, the standardized data of the residual terms should first be filtered. For example, a Savitzky-Golay filter can be used to filter (smooth) the standardized data of the residual terms to reduce the interference of large-amplitude noise. Then, the filtered standardized data is input into the corresponding single-stage LSTM regression model for training and prediction to obtain the prediction results of the residual terms. .

[0058] The Savitzky-Golley filter is a smoothing / denoising algorithm that performs polynomial least squares fitting within a sliding window. The core idea is to treat the data points within the window as "local samples," fit these points with a low-order polynomial, and then use the fitted value as the output of the center point. As the sliding window advances point by point, the entire curve is "polynomial smoothed" without losing peak height and slope information (this is the biggest difference from moving average).

[0059] The process of filtering the standardized data of the residual terms using the Savitzky-Golay filter specifically includes: First, consider the data points. Based on data time sequence position Set a filter window centered on the center, and set the window length. (Must be odd), the order of the polynomial is To avoid overfitting, set .

[0060] Then, perform least-squares fitting on the data points within the window: ; Among them, the summation variable , Therefore Data points within the center window These are the polynomial coefficients.

[0061] Once the fitting is complete, at the center point The fitted value at that point is used as the data point at the center of the window. Filtered output: .

[0062] As an equivalent linear operation, the output of the entire window can be represented as a fixed-weight vector: ; in, The coefficient combination is obtained through least squares solution, and depends on... and .

[0063] The process of training and predicting the standardized data of the residual terms using a single-level LSTM regression model is similar to the process of training and predicting the standardized data of the intrinsic mode functions using a single-level LSTM regression model, and will not be repeated here.

[0064] Finally, in step 160, The prediction results of each intrinsic mode function and the prediction results of the residual term are added together at the same time nodes of the time series to obtain the final power grid load prediction result: .

[0065] Evaluation metrics can also be designed to assess the power grid load forecasting results obtained by adaptive VMD-RF-LSTM-based load forecasting methods or methods utilizing such methods. These evaluation metrics mainly include: (1) Root Mean Squared Error (RMSE): ; in, It is the first The true value of each sample It is the first Prediction results for each sample It represents the total sample size; the square root of the squared mean of the prediction error "amplifies and punishes" large errors, making it susceptible to outliers; the smaller the value, the better, with 0 indicating a perfect fit.

[0066] (2) Mean Absolute Percentage Error (MAPE): ; The mean relative error percentage (MREP) represents the average percentage deviation of the predicted value from the actual value. It can be understood as the average magnitude of the prediction error. When the value is close to or equal to 0, the indicator diverges, distorting the entire MAPE indicator.

[0067] (3) Symmetric Mean Absolute Percentage Error (SMAPE): ; Change the denominator to This solves the problems of MAPE's sensitivity to zero and asymmetry, making it more robust and fairer.

[0068] (4) Coefficient of Determination R²: ; in, It is all true values The average value; The volatility ratio is used to explain the increase in the relative value of the benchmark model (mean), and the closer it is to 1, the better.

[0069] This study employs a multi-metric approach to evaluate method performance, going beyond simply using prediction error RMSE. In addition to traditional error metrics, it introduces evaluation metrics with better symmetry and robustness to outliers to comprehensively characterize the model's performance across different aspects. Multi-scale evaluation combines multiple metrics to comprehensively characterize model performance: RMSE directly reflects the magnitude of absolute error; MAPE provides a scale-independent comparison of relative error; SMAPE outperforms MAPE in terms of symmetry and robustness to extreme values; and the coefficient of determination... This indicates the model's ability to explain data variation, facilitating comparisons of the fit quality of different models. By combining these multi-scale metrics, predictive performance can be evaluated more comprehensively in different contexts, avoiding the misleading effects of relying on a single evaluation metric.

[0070] In summary, compared with directly inputting historical sequences into a general single model or existing technologies using hybrid models, this invention provides a systematic improvement in time series processing and exogenous information integration for the power grid load forecasting method based on adaptive VMD-RF-LSTM in the foregoing embodiments, in order to enhance the accuracy and stability of prediction. By constructing a power grid load forecasting method framework based on adaptive VMD-RF-LSTM, adaptive variational mode decomposition is used to decompose the complex actual power grid load time series into several harmonic components, and prediction is performed at the subsequence level. Subsequently, the residuals obtained from the decomposition are filtered and then predicted again, focusing on the multi-scale characteristics and robustness of the actual power grid load time series.

[0071] In the method steps of the foregoing embodiments, to further enhance the sensitivity of the prediction model to time-series features and rhythms, this invention introduces exogenous feature generation and designs corresponding time features in conjunction with industry practices. The initial intention of introducing these features is to more accurately capture the fluctuation patterns and activity levels of the power grid load sequence at different time periods, thereby improving the model's expressive power, robustness, and adaptability to time-series rhythms.

[0072] This invention also innovatively introduces VMD into the signal processing of actual power grid load time series during the data processing stage, decomposing complex signals into multiple harmonic components, and outputting the optimal VMD parameter combination, i.e., bandwidth constraint parameters, through an automatic optimization task of decomposition parameters. With the number of intrinsic mode functions This approach aims to provide the "cleanest" subsequences for subsequent prediction, noise reduction, or feature engineering. Furthermore, combining empirically generated exogenous features manually enhances the model's understanding of data structure and evolution patterns. Regarding data standardization, this invention departs from common max-scaling or min-max standardization, instead employing Median-Mad standardization to obtain robust estimates of center and dispersion. Additionally, the asinh transform is used to linearize the small-number region and logarithmically compress the large-number region, thereby enhancing robustness against outliers and improving model stability.

[0073] For the residual signal obtained from VMD decomposition, which often contains significant noise components, exhibiting high variance and strong volatility, this invention employs a single-stage LSTM for independent prediction of each sub-sequence. For the high-variance, highly volatile residual signal, it is first smoothed using a Savitzky-Golay filter to reduce interference from large-amplitude noise; then, a single-stage LSTM prediction is performed on the smoothed residual signal. Finally, the prediction results of each sub-sequence are added to the prediction results of the corresponding residual signal to form the final overall prediction output.

[0074] In a preferred embodiment, to overcome the drawback of error accumulation in long-term multi-step forecasting, an iterative optimization mechanism based on external variable forecasting feedback is added. This mechanism is used to apply the complete VMD-RF-LSTM forecasting process constructed in steps 130-160 to any other multi-source time series data (e.g., actual grid load) other than the actual grid load. Figure 4The external variables corresponding to the updated day-ahead electricity price data (as shown in the feedback process) are used to obtain their predicted data (predicted values). These predicted data are then used to update a portion of the original multi-source time series data (e.g., the time series data for the next 10 days). Using the updated multi-source time series data, the process returns to step 120 for exogenous feature generation, and steps 130-160 of the main path are executed to obtain the final grid load forecast result. This mechanism significantly improves the accuracy and adaptability of load forecasting by filtering external variables strongly correlated with grid load and feeding their future predicted values ​​back to the input data, forming a rolling closed loop of "prediction-update-re-prediction".

[0075] The iterative optimization mechanism based on external variable prediction feedback specifically includes: Step 1: Calculate the correlation index between the external variables and the corresponding variables of the actual power grid load in the original multi-source time series data; select the external variables whose correlation index is greater than the preset correlation threshold as strongly correlated external variables; the external variables refer to the corresponding variables of any other multi-source time series data besides the actual power grid load; the coherence index is the Pearson correlation coefficient or mutual information value.

[0076] In one embodiment, the coherence index is the Pearson correlation coefficient, and the preset correlation threshold is 0.6.

[0077] Step 2: For each strongly correlated external variable, perform mode decomposition according to the adaptive variational mode decomposition method in step 130 to obtain the corresponding intrinsic mode function and residual term; then, according to the process of steps 140 to 160, obtain the predicted value of the strongly correlated external variable in the future preset time period.

[0078] The historical time series of each strongly correlated external variable is modally decomposed according to the adaptive variational mode decomposition method in step 130, resulting in... There are one intrinsic mode function and one residual term. The number of modes is optimized using a sparrow search algorithm during the decomposition process. and penalty factor To minimize the envelope entropy.

[0079] Subsequently, steps 140 to 160 are executed sequentially: Median-MAD standardization and inverse hyperbolic sine transform are performed on each component, and the corresponding LSTM regression model is constructed for training and prediction. The prediction results of each component are then reconstructed after inverse transformation to obtain the predicted value sequence of the external variable within a preset future time period (e.g., the next 10 days). .

[0080] Step 3: Using the predicted values ​​of the strongly correlated external variables, update the time series data of the corresponding strongly correlated external variables in the original multi-source time series dataset for the preset time period, to obtain the updated multi-source time series dataset.

[0081] In one embodiment, the preset time period is the last 10 days.

[0082] Step 4: Based on the updated multi-source time series dataset, return to step 120 and sequentially execute steps 120-160, including regenerating exogenous features, performing adaptive variational mode decomposition, standardization and nonlinear transformation, long short-term memory network regression model prediction, and summing and reconstructing the prediction results, to obtain the updated power grid load prediction results.

[0083] Step 5: Repeat steps 2 to 4 until the preset iteration termination condition is met, thereby realizing the rolling optimization prediction of power grid load.

[0084] The preset iteration termination condition refers to satisfying at least one of the following conditions: The average absolute percentage error between the power grid load forecast results obtained from two consecutive iterations is less than the convergence threshold. ; The maximum number of iterations has been reached. .

[0085] In one embodiment, the convergence threshold The maximum number of iterations .

[0086] This invention upgrades the traditional open-loop forecasting system to a closed-loop adaptive system by introducing an iterative optimization mechanism based on external variable prediction feedback, thereby significantly improving the performance of power grid load forecasting. After the initial forecast, this mechanism filters out external variables strongly correlated with the load (such as...). Figure 4The framework adds a feedback path for day-ahead electricity prices (as shown in the diagram) and obtains their future forecasts. This updates the original dataset and re-forecasts load, forming a rolling closed loop of multiple iterative optimizations. First, by injecting the true trend information of external variables into subsequent forecasts, it effectively suppresses error accumulation in long-term multi-step forecasts, improving the accuracy of medium- to long-term forecasts. Second, this mechanism can uncover the deep coupling relationship between load and external variables, enabling the model to dynamically adapt to the impact of future changes in external conditions on load, outputting forecast results that better conform to actual physical and economic laws. Third, the iterative feedback process is equivalent to data augmentation based on model cognition, enhancing the model's robustness to data distribution shifts and extreme cases, achieving adaptive adjustment under rolling optimization. Finally, this framework coordinates load forecasting with key external variable forecasting, obtaining high-precision load results while also providing predicted values ​​for external variables, offering more comprehensive data support for power system dispatch decisions.

[0087] Furthermore, this invention verifies the beneficial effects of the proposed power grid load forecasting method based on adaptive VMD-RF-LSTM through specific experiments.

[0088] When using different model methods to achieve power grid load forecasting, Random Forest (RF) and XGBoost from traditional machine learning were selected as control models for baseline comparison. Meanwhile, a neural network with maximum normalization that does not introduce handcrafted exogenous features was chosen as a pure reference baseline.

[0089] Specifically, the neural network method models selected for data training and prediction (except for VMD-LSTM, all baselines use Min-Max normalization and do not contain manually set exogenous features) for comparison with the method model of this invention include: (1) A single-layer LSTM regression model with parameters consistent with the aforementioned single-level LSTM version to ensure controllability of the comparison; (2) Multi-CNN extracts local temporal features through layer-by-layer convolution; (3) CNN-LSTM uses convolutional layers for feature extraction and then connects them to LSTM for temporal modeling to combine local patterns and long-term dependencies; (4) VMD-LSTM: To highlight the effect of the residual filtering added in this invention, VMD-LSTM without residual filtering will be used as the baseline model for comparison. Its normalization and generated exogenous features are consistent with VMD-RF-LSTM.

[0090] The VMD-RF-LSTM model, which uses Median-Mad normalization and combines manually set exogenous features, is employed as the core comparison model. Four evaluation metrics—RMSE, MAPE, SMAPE, and... The model was compared with a baseline model on a dataset obtained from the power grid system of country F. The specific results are as follows: Figures 5-8 As shown. Figures 5-8 This diagram illustrates the comparison of results obtained using different models trained on a four-year power system dataset from country F, evaluating RMSE, MAPE, SMAPE, and R². The results show that VMD-RF-LSTM outperforms all control models in overall prediction accuracy, error distribution, and ability to capture data fluctuation trends, demonstrating stronger robustness and better fitting quality. Specifically, this includes a significant reduction in errors on RMSE, MAPE, and SMAPE, and... The explanatory power is significantly improved; in addition, the model maintains good stability and generalization ability even in datasets with different exogenous features, verifying the robustness and applicability of the proposed method across data.

[0091] In one embodiment, the present invention provides a power grid load forecasting device based on adaptive VMD-RF-LSTM, the device comprising: The first module is used to input multi-source time series data of the power system and perform preprocessing; the multi-source time series data includes at least actual generation data, dispatch plan boundary data, actual grid load and day-ahead electricity price; the preprocessed multi-source time series data is divided into datasets based on the same time sequence according to a preset ratio: training set, validation set and test set; The second module is used to construct features from the multi-source time series data after the dataset is divided, and to generate exogenous features for the training set, validation set or test set respectively; the exogenous features include at least multi-dimensional time features and dataset-inherent parameter features; the multi-source time series data other than the actual power grid load is concatenated with the corresponding exogenous features column by column to form an input data matrix; The third module is used to adaptively decompose the preprocessed actual grid load into... Subsequences, including Each intrinsic mode function and one residual term; each subsequence contains a dataset partition with the same time sequence as the actual grid load before decomposition; The fourth module is used to perform standardization and nonlinear transformation on each column of the input data matrix and each subsequence to obtain the corresponding standardized data. The fifth module is used to construct and train a single-level long short-term memory network regression model for each subsequence based on the standardized data of the training set. If the subsequence is a residual term, it is first filtered to suppress noise interference. The corresponding trained single-level long short-term memory network is used to predict the intrinsic mode function or the filtered residual term. After inverse standardization and inverse linear transformation, the prediction result of each subsequence is obtained. The sixth module is used to... The prediction results of each subsequence are summed at the same time nodes of the time series to obtain the final power grid load prediction result.

[0092] On the other hand, the present invention provides a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the power grid load forecasting method based on adaptive VMD-RF-LSTM provided in any of the above embodiments. The computer device may be a server. The computer device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The network interface of the computer device is used for communication with external terminals via a network connection.

[0093] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the power grid load forecasting method based on adaptive VMD-RF-LSTM provided in any of the above embodiments.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0095] Matters not covered in this invention are common knowledge. The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered to be within the scope of this specification.

[0096] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A power grid load forecasting method based on adaptive VMD-RF-LSTM, characterized in that, include: Step 110: Input multi-source time series data of the power system and perform preprocessing; The multi-source time series data includes at least actual generation data, dispatch plan boundary data, actual grid load and day-ahead electricity price; The preprocessed multi-source time series data are divided into training set, validation set and test set according to a preset ratio based on the same time series. Step 120: Construct features for the multi-source time series data after the dataset has been divided, and generate exogenous features for the training set, validation set, or test set respectively; the exogenous features include at least multi-dimensional time features and dataset-inherent parameter features; Multi-source time series data other than the actual power grid load are concatenated with corresponding exogenous features column-wise to form an input data matrix; Step 130: Using adaptive variational mode decomposition, the preprocessed actual power grid load is adaptively decomposed into... Subsequences, including Each intrinsic mode function and one residual term; each subsequence contains a dataset partition with the same time sequence as the actual grid load before decomposition; Step 140: Perform standardization and nonlinear transformation on each column of the input data matrix and each subsequence to obtain the corresponding standardized data; Step 150: Construct and train a single-level long short-term memory network regression model for each subsequence based on the standardized data of the training set; if the subsequence is a residual term, filter processing is performed first to suppress noise interference; use the corresponding trained single-level long short-term memory network to predict the intrinsic mode function or the filtered residual term, and then obtain the prediction result for each subsequence after inverse standardization and inverse linear transformation. Step 160, will The prediction results of each subsequence are summed at the same time nodes of the time series to obtain the final power grid load prediction result.

2. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 1, characterized in that, The actual data on the power generation side includes at least: biomass power generation, fossil lignite / lignite power generation, fossil gas power generation, fossil hard coal power generation, pumped storage power consumption, runoff and small reservoir hydropower generation, reservoir hydropower generation, nuclear power generation, conventional energy power generation not otherwise listed, solar power generation, waste power generation, onshore wind power, and renewable energy power generation not otherwise listed. The boundary data of the scheduling plan includes at least: day-ahead solar power forecast, day-ahead onshore wind power forecast, and total load forecast.

3. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 1, characterized in that, Preprocessing of any of the multi-source time series data includes at least missing value imputation and duplicate value processing; The missing value filling includes: if the time node When there are missing values, Take the previous time node With the next time node average As a time node Data filling at the location: ； It represents the total number of time points contained in the time series; The duplicate value processing includes: if time node Multiple data points exist at the same location: , ; If the number of duplicate data points is the number of data points, then the average of the duplicate data points is taken as the time node. Data at: 。 4. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 1, characterized in that, In step 120, the exogenous features include at least multi-dimensional temporal features and features with parameters inherent in the dataset. The multi-dimensional time features include at least: hour, day of the week, month, and business hours.

5. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 4, characterized in that, Step 130 includes: The pre-processed actual grid load The adaptive variational mode decomposition process can be formally described as an optimization problem: ， ； in, It is the first eigenmode functions The corresponding center frequency, , For unit impact function, This represents the convolution operation. It is the imaginary unit. It calculates the gradient with respect to time. Represents the square of the L2 norm; In parameter space The algorithm employs a sparrow search algorithm for global optimization. This algorithm simulates the foraging and anti-predation behavior of sparrows, efficiently exploring and developing within the parameter space through role updates of discoverer, follower, and watchdog, ultimately finding the globally optimal or near-optimal algorithm that minimizes the envelope entropy. ; Reprocess the actual grid load after preprocessing Perform an adaptive variational mode decomposition once, and finally obtain eigenmode functions and a residual term: 。 6. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 5, characterized in that, Step 140 includes: Median absolute deviation standardization is used to normalize the time series of each column of the input data matrix or each subsequence of the entire training set. Calculate the median of the whole. and absolute median ; Use the median and Time series Standardize: ； Using the inverse hyperbolic sine transform, for Standardized value The standardized data is obtained by transformation: 。 7. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 6, characterized in that, In step 150, if the subsequence is an intrinsic mode function, then the standardized data is input into the single-level long short-term memory network regression model corresponding to the subsequence, the model is trained using the standardized data of the training set corresponding to the subsequence, and the standardized data of the validation / test set corresponding to the subsequence is input into the trained single-level long short-term memory network regression model for prediction, to obtain the prediction result of the intrinsic mode function, including: For each subsequence, given the standardized input data, the 100 Long Short-Term Memory (LSTM) network units in the corresponding single-level LSM regression model scan the entire input time series with `return_sequences=True`. The model outputs a 100-dimensional hidden state at each time step in real time, capturing both local details and long-range dependencies simultaneously. The output shape is... ,in, It refers to time steps. `return_sequences` is a boolean parameter representing the returned sequence, used to control whether the output returned by the Long Short-Term Memory network layer is the hidden state of the last time step. This is the batch size, representing the number of samples input into the network at one time during model training; Using Flatten layer Tensor flattening The global context vector compresses time × features into a single long feature; The long features are input into the first fully connected layer Dense(200, ReLU) to learn high-level feature combinations in the spatial dimension; the first fully connected layer Dense has 200 neurons and uses the ReLU activation function to introduce non-linearity into the neural network; A high-dimensional nonlinear mapping is achieved by applying a 10% dropout to 200 ReLU neurons using a regularization layer Dropout (0.1). Dropout represents random deactivation, meaning that in each iteration of the training process, a subset of neurons in the network are randomly dropped; 0.1 indicates a dropout rate of 10%. The output layer Dense(1, linear) outputs a single scalar prediction to produce the predicted value; where linear is a linear activation function. The predicted values ​​output by the output layer are subjected to inverse transformation and inverse normalization to obtain the prediction results of the intrinsic mode functions corresponding to a subsequence. Repeat the above process to obtain the prediction results of all intrinsic mode functions, denoted as . .

8. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 7, characterized in that, In step 150, if the subsequence is a residual term, then the prediction result of the residual term is obtained by using variance evaluation, residual term filtering, and the single-level long short-term memory network regression model corresponding to the residual term, including: Calculate the variance (VAR) of the residual term; If the variance VAR is not greater than a given threshold The standardized data of the residual terms are directly input into the corresponding single-level long short-term memory network regression model for training and prediction, and the prediction results of the residual terms are obtained. If the variance VAR is greater than a given threshold The Savitzky-Golay filter is used to filter the standardized data of the residual terms to reduce noise interference in the standardized data. The filtered standardized data is then input into the single-level long short-term memory network regression model corresponding to the residual terms for training and prediction to obtain the prediction results of the residual terms.

9. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to claim 8, characterized in that, It also includes an iterative optimization step based on feedback prediction from external variables: Step 1: Calculate the correlation index between external variables and corresponding variables of actual power grid load in the original multi-source time series data; select external variables whose correlation index is greater than a preset correlation threshold as strongly correlated external variables; the external variable refers to the corresponding variable of any other multi-source time series data besides the actual power grid load; the coherence index is the Pearson correlation coefficient or mutual information value. Step 2: For each strongly correlated external variable, perform mode decomposition according to the adaptive variational mode decomposition method in step 130 to obtain the corresponding intrinsic mode function and residual term; then, according to the process of steps 140 to 160, obtain the predicted value of the strongly correlated external variable in the future preset time period. Step 3: Using the predicted values ​​of the strongly correlated external variables, update the time series data of the corresponding strongly correlated external variables in the original multi-source time series dataset during the preset time period to obtain the updated multi-source time series dataset. Step 4: Based on the updated multi-source time series dataset, return to step 120, and execute steps 120 to 160 sequentially to obtain the updated power grid load forecast results; Step 5: Repeat steps 2 to 4 until the preset iteration termination condition is met, thereby realizing the rolling optimization prediction of power grid load.

10. The power grid load forecasting method based on adaptive VMD-RF-LSTM according to any one of claims 1-9, characterized in that, It also includes design evaluation indicators to evaluate the method or the power grid load forecasting results obtained using the method; The evaluation indicators include at least: Root Mean Square Error (RMSE): ； in, It is the first The true value of each sample It is the first Prediction results for each sample It is the total sample size; Mean Absolute Percentage Error (MAPE): ； Symmetric Mean Absolute Percentage Error (SMAPE): ； Coefficient of determination R²: ； in, It is all real values The average value.

Images