Steel enterprise load prediction method and system based on rolling variational modal decomposition

Abstract

This invention discloses a method and system for load forecasting in steel enterprises based on rolling variational mode decomposition, belonging to the field of power load forecasting technology. The method includes: collecting load and external variable data; using multiple methods to jointly identify and correct outliers; filling missing values ​​with cubic spline interpolation; adaptively determining the window length based on the autocorrelation function; optimizing the variational mode decomposition parameters using a particle swarm optimization algorithm; performing rolling decomposition on the load sequence to obtain intrinsic mode function components; calculating the energy proportion and sample entropy of each component; constructing a weighted screening index to retain key components; using an ARIMA model to predict low-frequency trend components and an attention-time convolutional network to predict mid-to-high-frequency fluctuation components; constructing a weight learning module, inputting a feature vector containing multi-dimensional features, generating fusion weights for each component, and weighted summing to obtain the final predicted value. This invention avoids future information leakage and achieves frequency-domain differentiated modeling and adaptive weight fusion.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and more specifically, to a method and system for load forecasting of steel enterprises based on rolling variational mode decomposition. Background Technology

[0002] As a typical high-energy-consuming and high-emission industrial sector, steel enterprises rely heavily on electricity for their production processes, encompassing continuous operations across multiple stages, including ironmaking, steelmaking, continuous casting, and rolling. Influenced by the dynamic nature of production processes, frequent equipment start-ups and shutdowns, order fluctuations, and external environmental factors (such as temperature and humidity), their electricity load exhibits significant time-varying characteristics, multi-scale fluctuations, nonlinear evolution, and non-stationary statistical properties. Specifically, this manifests as frequent short-term load abrupt changes (such as the impact load during electric arc furnace smelting), a clear superposition of daily and weekly cycles, the coexistence of seasonal trends and random disturbances, and significant heteroscedasticity and long memory. Against this backdrop, achieving high-precision and robust short- to ultra-short-term load forecasting is not only directly related to the refined energy management within enterprises, the optimization of peak-valley electricity price arbitrage, and the improvement of demand-side response capabilities, but also has significant technical support value and strategic significance for the safe and stable operation of regional power grids, the scientific allocation of spinning reserve capacity, the coordinated improvement of renewable energy absorption capacity, and the effective implementation of the government's "dual carbon" target policy of "dual control" of total energy consumption and intensity.

[0003] Overcoming the limitations of single models, the "decomposition-prediction-reconstruction" paradigm has become a research hotspot in recent years. This paradigm follows the principle of "simplifying complexity and dividing and conquering," first decoupling the original complex load sequence into several physically meaningful sub-components (such as trend terms, periodic terms, and transient disturbance terms), then customizing the model for the characteristics of each component, and finally weighted and fused to recover the prediction result. Among these, variational mode decomposition (VMD), due to its adaptive bandwidth constraint mechanism based on variational principles, possesses excellent anti-modal aliasing capabilities, high spectral resolution, and strong stability of decomposition results. It has gradually replaced empirical methods such as EMD / EEMD and become the most mainstream preprocessing tool in the DPR framework. However, existing VMD-based forecasting methods still face two key technical bottlenecks. First, the problem of future information leakage is prominent. Second, the component modeling strategies are highly homogeneous and fail to reflect multi-scale heterogeneity. In summary, the next-generation intelligent model for load forecasting in steel enterprises urgently needs to achieve synergistic breakthroughs in two dimensions: ensuring temporal causality and driving modeling based on component characteristics. This will promote the evolution of forecasting technology from "usable" to an industrial-grade application paradigm that is "reliable, interpretable, and deployable". Summary of the Invention

[0004] This invention aims to solve the problems of future information leakage and single modeling strategy in the existing technology, and provides a method and system for load forecasting of steel enterprises based on rolling variational mode decomposition.

[0005] Firstly, a load forecasting method for steel enterprises based on rolling variational mode decomposition includes:

[0006] S1. Collect historical load data and related external variable data of steel enterprises, use multiple methods to jointly identify outliers and correct them, and use cubic spline interpolation to fill in missing data.

[0007] S2. The rolling window length is adaptively determined based on the autocorrelation function of the sequence. The particle swarm optimization algorithm is used to optimize the number of modes and the penalty factor of variational mode decomposition. The preprocessed load sequence is then subjected to rolling decomposition to obtain several intrinsic mode function components.

[0008] S3. Calculate the energy proportion and sample entropy of each modal component, construct a weighted screening index, and retain modal components whose index values ​​exceed the preset threshold.

[0009] S4. The selected low-frequency trend components are predicted using a seasonal autoregressive integral moving average model, and the medium- and high-frequency fluctuation components are predicted using a temporal convolutional network with an attention mechanism. The hyperparameters of the temporal convolutional network are adaptively optimized using the dung beetle optimization algorithm.

[0010] S5. Construct a weight learning module. Input the feature vector containing the predicted values ​​of each component, the width of the prediction confidence interval, local fluctuation characteristics, and external variables. Dynamically generate the fusion weights of each component through a multilayer perceptron. The weighted sum of the preliminary predicted values ​​of each component is used to obtain the final load prediction.

[0011] Optionally, the method of using multiple methods to jointly determine outliers in S1 specifically includes: simultaneously using the interquartile range method, the moving window method, and the isolated forest algorithm to determine each data point; if at least two methods determine it as an outlier, it is marked as an outlier point; for isolated outliers, a weighted average of the preceding and following normal values ​​is used instead; for continuous outlier segments, cubic spline interpolation is used for correction, and the power value is constrained to be non-negative.

[0012] Optionally, the adaptive method for determining the scroll window length L in S2 is as follows:

[0013] Calculate the autocorrelation function of the load sequence, take the lag order corresponding to the first peak of the autocorrelation function as the base period T, and set the window length. , where k is an integer, and the optimal candidate is selected from the candidate set of k by the particle swarm optimization algorithm. The optimization objective is to minimize the weighted sum of sample entropy and reconstruction error.

[0014] Optionally, the weighted screening index described in S3 The formula is as follows:

[0015]

[0016] in, Let i be the energy of the i-th modal component. Its sample entropy, , These are the weighting coefficients.

[0017] Optionally, in the temporal convolutional network with an attention mechanism described in S4, the attention mechanism uses scaled dot product attention, and its calculation formula is as follows:

[0018]

[0019] in, These are query, key, and value matrices, respectively. The dimension of the key; multi-head attention is a linear transformation that concatenates the results of multiple single-head attention.

[0020] Optionally, the feature vector input to the weight learning module in S5 specifically includes:

[0021] External variables at the current moment include temperature, humidity, rebar futures prices, and iron ore futures prices; the confidence interval width of the predicted values ​​for each component, obtained through quantile regression to obtain a 90% confidence interval; local fluctuation characteristics: standard deviation, kurtosis, and skewness of load values ​​over the past N moments, where N is 24; historical prediction errors for each component: mean absolute error of each component over the past 24 hours; time encoding characteristics: sine and cosine encoding of hours, days of the week, and holidays; normalized value of the current load; predicted values ​​of each component at the current moment; energy proportion of each component; interaction characteristics: product of load and temperature, and product of load and futures prices;

[0022] Optionally, the feature vector is input into a multilayer perceptron after standardization. The multilayer perceptron contains two hidden layers with 32 and 16 neurons respectively, and the activation function is ReLU. The output layer uses the Softmax activation function to generate weights.

[0023] Optionally, the prediction method further includes S6 concept drift detection and adaptive update, constructing a prediction performance monitor, calculating prediction error statistics within the sliding window in real time, and setting multi-level early warning thresholds; when a performance degradation is detected, different update strategies are triggered according to the degree of degradation, including online incremental update, full retraining, or model fusion.

[0024] Optionally, the concept drift detection and adaptive update step specifically includes:

[0025] S601. Maintain a sliding window of length N, record the absolute error of the prediction at each time step; calculate the mean absolute error and standard deviation of the error within the window.

[0026] S602, Set the basic threshold, the severity threshold, and the fluctuation threshold;

[0027] S603. When the average absolute error of the window is greater than or equal to the base threshold or the standard deviation of the error within the window is greater than the preset fluctuation threshold, it is determined to be a slight concept drift.

[0028] S604. When the basic threshold of a window is greater than or equal to the severe threshold, or when multiple consecutive windows meet the criteria for mild concept drift, it is determined to be severe concept drift.

[0029] S605. When a mild concept drift is detected, an online incremental update is triggered. The weight network is fine-tuned with a small amount of gradient descent using the data in the most recent window, and the new data is added to the training cache.

[0030] S606. When a severe concept drift is identified, a full retraining is triggered. The rolling VMD decomposition, hierarchical model training, and weight network training are re-executed using data from the most recent period, and the online model is replaced.

[0031] Secondly, a load forecasting system for steel enterprises based on rolling variational mode decomposition includes:

[0032] The data acquisition and preprocessing module is used to collect historical load data and related external variable data of steel enterprises, use multiple methods to jointly identify and correct outliers, and use cubic spline interpolation to fill in missing data.

[0033] The rolling VMD decomposition module is used to adaptively determine the rolling window length based on the autocorrelation function of the sequence. It uses the particle swarm optimization algorithm to optimize the number of modes and the penalty factor of variational mode decomposition, and performs rolling decomposition on the preprocessed load sequence to obtain several intrinsic mode function components.

[0034] The component filtering module is used to calculate the energy proportion and sample entropy of each modal component, construct a weighted filtering index, and retain modal components whose index values ​​exceed a preset threshold.

[0035] The hierarchical prediction module is used to predict the selected low-frequency trend components using a seasonal autoregressive integral moving average model, and to predict the medium- and high-frequency fluctuation components using a temporal convolutional network with an attention mechanism. The hyperparameters of the temporal convolutional network are adaptively optimized using the dung beetle optimization algorithm.

[0036] The dynamic weight fusion module is used to build the weight learning module. The input includes the feature vector containing the predicted values ​​of each component, the prediction confidence interval width, local fluctuation characteristics and external variables. The fusion weight of each component is dynamically generated through a multilayer perceptron. The preliminary predicted values ​​of each component are weighted and summed to obtain the final load forecast.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention addresses the high volatility and non-stationarity characteristics of load forecasting in steel enterprises by proposing a short-term load forecasting method based on rolling Virtual Machine Decomposition (VMD). The introduction of VMD ensures that the decomposition process uses only historical information, resolving the problem of future information leakage in traditional decomposition forecasting models and improving the model's operability in real-world environments. Simultaneously, the introduction of the Producer-Solver (PSO) algorithm enables adaptive optimization of decomposition parameters, reducing reliance on human experience. By employing targeted forecasting methods for load components at different frequency scales, the modeling advantages of various models are fully utilized, resulting in improved forecasting accuracy. Attached Figure Description

[0039] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0040] Figure 1 This is a flowchart of the load prediction method for steel enterprises based on rolling variational mode decomposition of the present invention.

[0041] Figure 2 This is a roadmap for the hierarchical model technology based on rolling VMD in this embodiment of the invention.

[0042] Figure 3 This is a power load curve of a steel company in an embodiment of the present invention.

[0043] Figure 4 This is a block diagram of the steel enterprise load prediction system based on rolling variational mode decomposition according to the present invention. Detailed Implementation

[0044] 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 a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0046] Example 1

[0047] One embodiment of this invention discloses a load forecasting method for steel enterprises based on rolling variational mode decomposition. The method introduces rolling VMD to ensure that the decomposition process uses only historical information, including:

[0048] S1. Collect historical load data and related external variable data of steel enterprises, use multiple methods to jointly identify outliers and correct them, and use cubic spline interpolation to fill in missing data.

[0049] Outliers were identified and corrected, and missing data were filled using cubic spline interpolation.

[0050] The method of using multiple methods to jointly identify and correct outliers is as follows:

[0051] For each data point X t Three methods are used to determine this simultaneously:

[0052] IQR method: Calculate the global quartiles Q1 and Q3, IQR = Q3 - Q1. If X t <Q1-1.5×IQR or X t >Q3+1.5×IQR, mark as an anomaly.

[0053] Moving window method: Window width W=24, calculate the mean μ within the window. w and standard deviation σ w If |X t −μ w |>3σ w Mark as an anomaly;

[0054] Isolation Forest: Set the pollution rate to 0.05, train the model, and mark anomalies with an anomaly score higher than 0.6.

[0055] If at least two methods mark it as an anomaly, then it is determined to be an anomaly.

[0056] The correction of outliers is as follows:

[0057] Isolated outliers: Take the three normal values ​​before and after the outlier, and replace them with a weighted average of the inverse time distance.

[0058] Continuous abnormal segments: Cubic spline interpolation is used, with 5 normal points before and after the abnormal segment as endpoints, to generate a smooth curve to replace all points in the abnormal segment, and the power value is constrained to be non-negative.

[0059] Missing data are filled using cubic spline interpolation, as detailed below:

[0060] For missing points, cubic spline interpolation is used, and the interpolation function is fitted using 10 normal points before and after the missing points to calculate the estimated value of the missing points.

[0061] S2. The rolling window length is adaptively determined based on the autocorrelation function of the sequence. The particle swarm optimization algorithm is used to optimize the number of modes and the penalty factor of variational mode decomposition. The preprocessed load sequence is then subjected to rolling decomposition to obtain several intrinsic mode function components.

[0062] The specific process for rolling VMD decomposition and parameter optimization is as follows:

[0063] Calculate the autocorrelation function (ACF) of the load series, τ = 0, 1, ..., 288. Take the lag order T corresponding to the first peak of the ACF (excluding τ = 0) as the base period. Set the candidate window length L = k × T, k ∈ {2, 3, 4, 5}. Optimize in the joint space of k, K, and α using PSO.

[0064] MD decomposes the signal f(t) into K modes uk(t) by constructing and solving a constrained variational problem:

[0065]

[0066] The constraints are:

[0067] Where k is the number of mode decompositions, uk and ωk are the kth mode component and center frequency, respectively; f(t) is the input original signal; δ(t) is the Dirac function; and ∂(t) is the first-order partial derivative at time t.

[0068] By introducing a quadratic penalty factor α and Lagrange multipliers λ(t), the problem is transformed into an unconstrained problem. The constrained variational problem is then solved as an unconstrained variational problem, as shown in the following formula:

[0069]

[0070] The optimal solution is obtained iteratively using the alternating direction multiplier method.

[0071] The constructed rolling VMD decomposition, which extracts IMF components from the end of the window, effectively avoids future information leakage. The steps are as follows:

[0072] Step 1, Window Size L Setting: Set the fixed window length L used to extract local data segments;

[0073] Step 2, Setting the Rolling Step Size s: Decompose stepwise, setting the step size to 1. The window advances one time step at a time, using only the data within the window for VMD decomposition to avoid future information leakage;

[0074] Step 3, Data Reconstruction: Reconstruct the time series data into a two-dimensional matrix with a window length of L and a rolling step of s, with dimensions (N−L−s+1,L), where N is the length of the time series data;

[0075] Step 4: Rolling decomposition: The reconstructed two-dimensional matrix is ​​decomposed row by row using VMD to obtain a three-dimensional array of decomposition results with dimensions (N−L−s+1,L,K), where K is the number of sub-modes.

[0076] Step 5: Extracting end data: Extract the decomposed values ​​from the end time of each window and splice them into a complete modality sequence in chronological order.

[0077] The complex hyperparameters of VMD affect the effectiveness of decomposition, and traditional methods struggle to efficiently find the globally optimal combination. This invention employs the PSO algorithm for optimization.

[0078] S3. Calculate the energy proportion and sample entropy of each modal component, construct a weighted screening index, and retain modal components whose index values ​​exceed the preset threshold.

[0079] In this embodiment, the specific operation process of this step is as follows:

[0080] The specific process is as follows:

[0081] The energy and sample entropy of each modality are calculated to construct a weighted screening index, as shown in the following formula:

[0082]

[0083] in, Let i be the energy of the i-th modal component. Its sample entropy, , Let be the weighting coefficient, satisfying + =1, reserved Components whose values ​​are greater than a preset threshold.

[0084] Leveraging the advantages of VMD in decomposing non-stationary signals and the ability of rolling windows to control data leakage, this study decomposes the load sequence of steel enterprises. Parameters are optimized using the PSO algorithm. Based on the optimal parameter decomposition, key components are extracted using the energy thresholding method. These components are treated as independent sequences and, together with exogenous variables, undergo feature engineering. They are then input into corresponding modules for prediction. The prediction results of each component are superimposed to reconstruct the power load prediction value. This approach preserves the independence of different frequency characteristics while improving prediction accuracy through parameter optimization and component selection.

[0085] S4. The selected low-frequency trend components are predicted using a seasonal autoregressive integral moving average model, and the medium- and high-frequency fluctuation components are predicted using a temporal convolutional network with an attention mechanism. The hyperparameters of the temporal convolutional network are adaptively optimized using the dung beetle optimization algorithm.

[0086] Please see Figure 2 The specific process is as follows:

[0087] Specifically, a temporal convolutional network with an attention mechanism is used to predict the mid-to-high frequency fluctuation components. The attention mechanism in this network employs scaled dot product attention, calculated using the following formula:

[0088]

[0089] in, These are query, key, and value matrices, respectively. Let T be the dimension of the key, and T denote the matrix transpose operation; multi-head attention is a linear transformation that concatenates the results of multiple single-head attention operations.

[0090] The hyperparameters of the temporal convolutional network are adaptively optimized using the Dung Beetle Optimization (DBO) algorithm.

[0091] Based on the frequency domain characteristics of variational mode decomposition and the structural features of time series, this invention adopts a hierarchical modeling strategy: the dominant component IMF1 is modeled using the ARIMA model, while the fluctuation component is modeled using the TCN model. IMF1 mainly reflects the long-term trend and slow-varying characteristics of the data. The ARIMA model can efficiently model the trend dependence after deperiodicization, and has the advantages of simple parameters and high computational efficiency. The fluctuation component, however, contains complex nonlinear dynamic features and short-term fluctuation patterns, requiring modeling using the temporal convolution mechanism and nonlinear mapping capabilities of the TCN. This frequency-domain modeling strategy ensures the stability of low-frequency trend modeling while fully leveraging the advantages of deep networks in complex fluctuation modeling.

[0092] S5. Construct a weight learning module. Input the feature vector containing the predicted values ​​of each component, the width of the prediction confidence interval, local fluctuation characteristics, and external variables. Dynamically generate the fusion weights of each component through a multilayer perceptron. The weighted sum of the preliminary predicted values ​​of each component is used to obtain the final load prediction.

[0093] In this embodiment, the input feature vector specifically includes:

[0094] External variables at the current moment include temperature, humidity, rebar futures prices, and iron ore futures prices; the confidence interval width of the predicted values ​​for each component, obtained through quantile regression to obtain a 90% confidence interval; local fluctuation characteristics: standard deviation, kurtosis, and skewness of load values ​​over the past N moments, where N is 24; historical prediction errors for each component: mean absolute error of each component over the past 24 hours; time encoding characteristics: sine and cosine encoding of hours, days of the week, and holidays; normalized value of the current load; predicted values ​​of each component at the current moment; energy proportion of each component; interaction characteristics: product of load and temperature, and product of load and futures prices;

[0095] The feature vectors are standardized and then input into a multilayer perceptron. The multilayer perceptron contains two hidden layers with 32 and 16 neurons respectively, and the activation function is ReLU. The output layer uses the Softmax activation function to generate weights.

[0096] The prediction method also includes S6 concept drift detection and adaptive update, constructing a prediction performance monitor, calculating prediction error statistics within the sliding window in real time, and setting multi-level early warning thresholds; when a performance degradation is detected, different update strategies are triggered according to the degree of degradation, including online incremental update, full retraining, or model fusion.

[0097] In this step, the concept drift detection and adaptive update steps are performed as follows:

[0098] S601. Maintain a sliding window of length N, record the absolute error of the prediction at each time step; calculate the mean absolute error and standard deviation of the error within the window.

[0099] In the online prediction system, the absolute prediction error is recorded in real time at every moment. The formula for calculating the mean absolute error within the window is as follows:

[0100]

[0101] in, Indicates the size of the sliding window. This represents the absolute error of the prediction at time t. This represents the actual load value at time t (the power data actually collected). This represents the model prediction value at time t;

[0102] The formula for calculating the standard deviation of error is as follows:

[0103]

[0104] S602, Set the basic threshold, the severity threshold, and the fluctuation threshold;

[0105] S603. When the mean absolute error of the window is greater than or equal to the base threshold (and less than the severe threshold) or the standard deviation of the error within the window is greater than the preset fluctuation threshold, it is determined to be a mild concept drift.

[0106] S604. When the basic threshold of a window is greater than or equal to the severe threshold, or when multiple consecutive windows meet the criteria for mild concept drift, it is determined to be severe concept drift.

[0107] S605. When a mild concept drift is detected, an online incremental update is triggered. The weight network is fine-tuned with a small amount of gradient descent using the data in the most recent window, and the new data is added to the training cache.

[0108] S606. When a severe concept drift is identified, a full retraining is triggered. The rolling VMD decomposition, hierarchical model training, and weight network training are re-executed using data from the most recent period, and the online model is replaced.

[0109] In practical application and implementation: This embodiment takes a steel company as an example, collecting actual data from 00:00 on July 3, 2024 to 00:00 on December 18, 2024 to verify the effectiveness of the layered modeling method based on PSO optimization and rolling window for load forecasting of steel companies. All numerical information is displayed after anonymization. The power load curve of the observed company is shown in Figure 3. The company's load curve exhibits obvious periodic fluctuations and intermittent operation patterns, repeatedly fluctuating at high and low load levels with large amplitude changes, and is characterized by a wide numerical range and large fluctuations.

[0110] The load characteristics of steel enterprises are complex. When selecting inputs for the forecasting model, the influence of historical loads must be considered, along with material prices, date types, and meteorological factors. The following variables were selected: actual sampled historical load data, with a sampling interval of 15 minutes; meteorological factors including temperature (°C) and humidity (%), collected from the China Meteorological Administration; and material price factors including steel futures prices and iron ore futures prices, represented by the closing prices (yuan / ton) of rebar futures and iron ore futures, collected from Steel Union data.

[0111] Unavoidable outliers in the raw load data were identified using a combination of IQR, moving window, and IsolationForest methods. Points identified as outliers by at least two of these methods were considered outliers. Isolated outliers were replaced with the average of preceding and following normal values, while consecutive outliers were corrected using linear interpolation, with the power constrained to be non-negative. Meteorological data, considering physical constraints, statistical anomalies, and seasonal factors, were replaced with linear interpolation to replace outliers. Approximately 5% of the metering data points were missing and were filled in using interpolation.

[0112] The rolling VMD method was used to decompose the original load data, and the PSO algorithm was used to adaptively select its parameters. The weighted sum of sample entropy and reconstruction error was used as the overall fitness function. With 15 particles and 30 iterations, the optimal parameter combination was obtained: number of decomposition modes K=8, equilibrium parameter α=1211, and window length of 400.

[0113] The correlation coefficient for reconstruction quality assessment reached 0.9901, indicating excellent decomposition results. Figure 4 shows the decomposition results of a portion of the sequence extracted from the selected time period. The energy distribution of each modal component is clearly hierarchical: IMF1, as the dominant component, accounts for 41.91% of the signal energy, with a mean of 6.61, reflecting the basic level characteristics and long-term trend of the load; IMF2 and IMF3 account for 29.96% and 17.83% of the energy, respectively, with mean values ​​close to zero, exhibiting pure fluctuation characteristics and carrying the main variation pattern of the load; IMF4-8 have gradually higher frequencies and their energy proportions decrease step by step, which verifies the effectiveness of VMD decomposition.

[0114] Based on the above decomposition results, IMF components are screened using energy thresholds. By calculating the variance contribution rate of each modal component, key components with a variance proportion exceeding a preset threshold of 1% are automatically identified and retained, effectively filtering out low-energy noise components and improving the signal-to-noise ratio and computational efficiency of subsequent modeling. Calculations show that IMF1, IMF2, and IMF3 are retained for subsequent modeling.

[0115] To fully explore the fluctuation characteristics of power load time series data and the frequency domain information of VMD decomposition, a multi-level feature engineering system was constructed, including time features, lag features, rolling statistical features, multi-timescale difference features, interaction features, and external factors such as meteorology. Simultaneously, by integrating knowledge from the power industry, business-oriented feature variables such as weekday-holiday patterns and overload production indicators were constructed. For the unique frequency characteristics of different modal components, a gradient boosting regression was used to calculate the feature importance score for each component individually. The top 50 most relevant features were selected from the entire feature set to generate a dedicated modeling dataset.

[0116] All numerical experiments were conducted on a 64-bit Windows operating system computer with 40 GB of memory and a Quadro RTX 4000 GPU. Algorithm programming was performed on Anaconda using Python 3.9 and the PyTorch deep learning framework.

[0117] The dataset is divided into two parts in a 6:2:2 ratio according to time sequence. Data from the 24 time steps before time t is input into the model to predict the enterprise's two-step ultra-short-term load.

[0118] To evaluate the performance of the model, the proposed model was compared with five models: three models without rolling VMD (ARIMA, TCN, and LSTM) and two models after rolling VMD. The models after rolling VMD were: (1) Rolling VMD-Unified-TCN model: the model that uses the TCN model to directly predict the original load sequence after enhancing the IMF components as features; (2) Rolling VMD-Unified-LSTM model: the model that uses the same strategy as model (1) but uses the LSTM architecture. Different hyperparameter tuning strategies were adopted for different models, and the same dataset, optimization method, loss function and regularization strategy were used to predict the same task.

[0119] Example 2

[0120] See Figure 4 One embodiment of this disclosure provides a load forecasting system for steel enterprises based on rolling variational mode decomposition, comprising:

[0121] The data acquisition and preprocessing module is used to collect historical load data and related external variable data of steel enterprises, use multiple methods to jointly identify and correct outliers, and use cubic spline interpolation to fill in missing data.

[0122] The rolling VMD decomposition module is used to adaptively determine the rolling window length based on the autocorrelation function of the sequence. It uses the particle swarm optimization algorithm to optimize the number of modes and the penalty factor of variational mode decomposition, and performs rolling decomposition on the preprocessed load sequence to obtain several intrinsic mode function components.

[0123] The component filtering module is used to calculate the energy proportion and sample entropy of each modal component, construct a weighted filtering index, and retain modal components whose index values ​​exceed a preset threshold.

[0124] The hierarchical prediction module is used to predict the selected low-frequency trend components using a seasonal autoregressive integral moving average model, and to predict the medium- and high-frequency fluctuation components using a temporal convolutional network with an attention mechanism. The hyperparameters of the temporal convolutional network are adaptively optimized using the dung beetle optimization algorithm.

[0125] The dynamic weight fusion module is used to build the weight learning module. The input includes the feature vector containing the predicted values ​​of each component, the prediction confidence interval width, local fluctuation characteristics and external variables. The fusion weight of each component is dynamically generated through a multilayer perceptron. The preliminary predicted values ​​of each component are weighted and summed to obtain the final load forecast.

[0126] This embodiment provides a load forecasting system for steel enterprises based on rolling variational mode decomposition. Through advanced data processing, signal decomposition, and intelligent prediction algorithms, it achieves high-precision prediction of the power load of complex steel enterprises.

[0127] The steel enterprise load forecasting system based on rolling variational mode decomposition according to embodiments of the present invention can correspond to the execution of the method described in the embodiments of the present invention, and the other operations and / or functions of each module / unit of the steel enterprise load forecasting system based on rolling variational mode decomposition are respectively for implementing Figure 1 For the sake of brevity, the corresponding processes of each method in the illustrated embodiments will not be described in detail here.

[0128] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

[0129] Many other changes and modifications can be made without departing from the concept and scope of this invention. It should be understood that this invention is not limited to the specific embodiments, and the scope of this invention is defined by the appended claims.

Claims

1. A load forecasting method for steel enterprises based on rolling variational mode decomposition, characterized in that, include: S1. Collect historical load data and related external variable data of steel enterprises, use multiple methods to jointly identify outliers and correct them, and use cubic spline interpolation to fill in missing data. S2. The rolling window length is adaptively determined based on the autocorrelation function of the sequence. The particle swarm optimization algorithm is used to optimize the number of modes and the penalty factor of variational mode decomposition. The preprocessed load sequence is then subjected to rolling decomposition to obtain several intrinsic mode function components. S3. Calculate the energy proportion and sample entropy of each modal component, construct a weighted screening index, and retain modal components whose index values ​​exceed the preset threshold. S4. The selected low-frequency trend components are predicted using a seasonal autoregressive integral moving average model, and the medium- and high-frequency fluctuation components are predicted using a temporal convolutional network with an attention mechanism. The hyperparameters of the temporal convolutional network are adaptively optimized using the dung beetle optimization algorithm. S5. Construct a weight learning module. Input the feature vector containing the predicted values ​​of each component, the width of the prediction confidence interval, local fluctuation characteristics, and external variables. Dynamically generate the fusion weights of each component through a multilayer perceptron. The weighted sum of the preliminary predicted values ​​of each component is used to obtain the final load prediction.

2. The method for load forecasting of steel enterprises based on rolling variational mode decomposition according to claim 1, characterized in that, The method of using multiple methods to jointly determine outliers as described in S1 specifically includes: simultaneously using the interquartile range method, the moving window method, and the isolated forest algorithm to determine each data point; if at least two methods determine it as an outlier, it is marked as an outlier point; for isolated outliers, a weighted average of the preceding and following normal values ​​is used instead; for continuous outlier segments, cubic spline interpolation is used for correction, and the power value is constrained to be non-negative.

3. The method for load forecasting of steel enterprises based on rolling variational mode decomposition according to claim 1, characterized in that, The adaptive method for determining the scroll window length L in S2 is as follows: Calculate the autocorrelation function of the load sequence, take the lag order corresponding to the first peak of the autocorrelation function as the base period T, and set the window length. , where k is an integer, and the optimal candidate is selected from the candidate set of k by the particle swarm optimization algorithm. The optimization objective is to minimize the weighted sum of sample entropy and reconstruction error.

4. The method for load forecasting of steel enterprises based on rolling variational mode decomposition according to claim 1, characterized in that, The weighted screening index described in S3 The formula is as follows: ； in, Let i be the energy of the i-th modal component. Its sample entropy, , These are the weighting coefficients.

5. The method for load forecasting of steel enterprises based on rolling variational mode decomposition according to claim 1, characterized in that, In the temporal convolutional network with an attention mechanism described in S4, the attention mechanism uses scaled dot product attention, and the calculation formula is as follows: ； in, These are query, key, and value matrices, respectively. The dimension of the key; multi-head attention is a linear transformation that concatenates the results of multiple single-head attention.

6. The method for load forecasting of steel enterprises based on rolling variational mode decomposition according to claim 1, characterized in that, The specific feature vector input to the weight learning module described in S5 includes: External variables at the current moment include temperature, humidity, rebar futures prices, and iron ore futures prices; the confidence interval width of the predicted values ​​for each component, obtained through quantile regression to obtain a 90% confidence interval; local fluctuation characteristics: standard deviation, kurtosis, and skewness of load values ​​over the past N moments, where N is 24; historical prediction errors for each component: mean absolute error of each component over the past 24 hours; time encoding characteristics: sine and cosine encoding of hours, days of the week, and holidays; normalized value of the current load; predicted values ​​of each component at the current moment; energy proportion of each component; interaction characteristics: product of load and temperature, and product of load and futures prices; The feature vectors are standardized and then input into a multilayer perceptron. The multilayer perceptron contains two hidden layers with 32 and 16 neurons respectively, and the activation function is ReLU. The output layer uses the Softmax activation function to generate weights.

7. The method for load forecasting of steel enterprises based on rolling variational mode decomposition according to claim 1, characterized in that, The prediction method also includes S6 concept drift detection and adaptive update, constructing a prediction performance monitor, calculating prediction error statistics within the sliding window in real time, and setting multi-level early warning thresholds; when a performance degradation is detected, different update strategies are triggered according to the degree of degradation, including online incremental update, full retraining, or model fusion.

8. The method for load forecasting of steel enterprises based on rolling variational mode decomposition according to claim 7, characterized in that, The concept drift detection and adaptive update step specifically includes the following process: S601. Maintain a sliding window of length N, record the absolute error of the prediction at each time step; calculate the mean absolute error and standard deviation of the error within the window. S602, Set the basic threshold, the severity threshold, and the fluctuation threshold; S603. When the average absolute error of the window is greater than or equal to the base threshold or the standard deviation of the error within the window is greater than the preset fluctuation threshold, it is determined to be a slight concept drift. S604. When the basic threshold of a window is greater than or equal to the severe threshold, or when multiple consecutive windows meet the criteria for mild concept drift, it is determined to be severe concept drift. S605. When a mild concept drift is detected, an online incremental update is triggered. The weight network is fine-tuned with a small amount of gradient descent using the data in the most recent window, and the new data is added to the training cache. S606. When a severe concept drift is identified, a full retraining is triggered. The rolling VMD decomposition, hierarchical model training, and weight network training are re-executed using data from the most recent period, and the online model is replaced.

9. A load forecasting system for steel enterprises based on rolling variational mode decomposition, characterized in that, include: The data acquisition and preprocessing module is used to collect historical load data and related external variable data of steel enterprises, use multiple methods to jointly identify and correct outliers, and use cubic spline interpolation to fill in missing data. The rolling VMD decomposition module is used to adaptively determine the rolling window length based on the autocorrelation function of the sequence. It uses the particle swarm optimization algorithm to optimize the number of modes and the penalty factor of variational mode decomposition, and performs rolling decomposition on the preprocessed load sequence to obtain several intrinsic mode function components. The component filtering module is used to calculate the energy proportion and sample entropy of each modal component, construct a weighted filtering index, and retain modal components whose index values ​​exceed a preset threshold. The hierarchical prediction module is used to predict the selected low-frequency trend components using a seasonal autoregressive integral moving average model, and to predict the medium- and high-frequency fluctuation components using a temporal convolutional network with an attention mechanism. The hyperparameters of the temporal convolutional network are adaptively optimized using the dung beetle optimization algorithm. The dynamic weight fusion module is used to build the weight learning module. The input includes the feature vector containing the predicted values ​​of each component, the prediction confidence interval width, local fluctuation characteristics and external variables. The fusion weight of each component is dynamically generated through a multilayer perceptron. The preliminary predicted values ​​of each component are weighted and summed to obtain the final load forecast.

Images