Method and system for predicting short-term power load of industrial park
By using segmented clustering and hybrid model optimization, and by optimizing the autoregressive integral moving average model with time-series autoencoders and artificial lemming algorithms, the problems of data scarcity and load curve complexity in short-term power load forecasting of industrial parks are solved, achieving high-precision and widely applicable load forecasting.
Patent Information
- Application Number
- CN202511763023.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-03-03
AI Technical Summary
Existing technologies are unable to effectively address the problems of complex load curve shapes caused by responding to peak-valley-flat electricity pricing, reduced historical data reference value due to rapid product upgrades, and lack of model training data due to incomplete data records, especially in short-term power load forecasting in industrial parks.
A method combining segmented clustering and hybrid model optimization is adopted. By extracting features from segmented load curves and time-series autoencoders, and combining artificial lemming algorithm to optimize autoregressive integral moving average model and gated cyclic unit, the linear and nonlinear dynamics of load sequence are captured to achieve high-precision prediction.
Under conditions where data volume and quality requirements are not high, adaptive prediction of load characteristics of industrial parks is achieved, improving the prediction capability and generalization of small samples. It is applicable to various differentiated parks and has high accuracy and strong generalization.
Smart Images
Figure CN121602344A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power load forecasting technology, specifically relating to a method and system for short-term power load forecasting in industrial parks. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] With the deepening of electricity market reform, time-of-use pricing policies (usually divided into peak, off-peak, and valley periods) have become an important means of guiding users to smooth out peak demand and optimize the allocation of electricity resources. Against this backdrop, enterprises in industrial parks generally proactively adjust their production plans based on electricity price signals to reduce electricity costs. For example, they shift energy-intensive production processes from peak and off-peak periods when electricity prices are higher to valley periods when prices are lower. This proactive load adjustment behavior based on economic incentives results in a significant and non-linear distortion of the daily electricity load curve in industrial parks compared to the natural load curve, making its fluctuation patterns more complex and greatly increasing the difficulty of accurate forecasting.
[0004] On the other hand, fierce market competition drives industrial enterprises to continuously upgrade their technology and iterate their products. The transformation of production lines, the introduction of new products, and the optimization of production processes all mean that their electricity consumption patterns and energy consumption characteristics will undergo fundamental changes. This change in load characteristics caused by the inherent updates in production activities renders historical electricity consumption data from earlier periods unreliable for predicting future load conditions. In other words, the correlation between historical data and future load weakens, and the predictive performance of traditional time-series forecasting models that rely on long-term, stable historical data will decline sharply in such scenarios.
[0005] Furthermore, many industrial parks, especially newly built ones or those with low levels of informatization, generally suffer from problems such as short historical load data accumulation periods, incomplete data records, or coarse sampling granularity. The scarcity of available data makes it easy for many advanced prediction models (such as deep learning models) that rely on large amounts of data for training to fall into the trap of overfitting or insufficient training, making it difficult to uncover effective prediction patterns. Summary of the Invention
[0006] To address the aforementioned problems, this invention proposes a method and system for short-term power load forecasting in industrial parks. This invention can adapt to rapid changes in load patterns and does not have high requirements for the quantity and quality of historical data. Through segmented clustering and hybrid model collaborative optimization, it solves the problems of load curve shape differences and model adaptability.
[0007] According to some embodiments, the present invention adopts the following technical solution: A method for short-term power load forecasting in industrial parks includes the following steps: Obtain load data from the industrial park, construct load curves, and segment and preprocess the load curves. Extract short-term features and variable coupling relationships for each load segment, and capture the sequential dependencies over a long time period. Reconstruct the original load sequence from the extracted feature vectors. Based on the feature vectors extracted from each load segment, the morphological similarity between each feature vector is calculated. Based on the calculated similarity, the load segments are clustered into K clusters. The parameters of the autoregressive integral moving average model are optimized using the artificial lemming algorithm. The optimized autoregressive integral moving average model is used to make an initial prediction of the load sequence to obtain the predicted value reflecting the linear component of the sequence. The residual sequence between the predicted value and the actual load value is calculated. The nonlinear dynamics and long-term dependencies in the sequence are captured by the gated cyclic unit to predict the future residual value. The predicted value, which reflects the linear component of the sequence, is added to the future residual value to obtain the final load forecast.
[0008] As an alternative implementation method, the process of segmenting the load curve includes cutting the complete daily electricity load curve into multiple independent load segments according to the time points published by the region where the industrial park is located.
[0009] As an alternative implementation method, the process of preprocessing the load curve includes: Savitzky-Golay filters are used to denoise each load segment while preserving local morphological features. Linear interpolation is used to unify load segments of different lengths to the number of points on the longest curve in the set, thus achieving data alignment. Min-Max normalization is used to scale the load values to the [0,1] interval for normalization, eliminating the influence of dimensions.
[0010] As an alternative implementation, the process of extracting short-term features and variable coupling relationships of each load segment and capturing sequential dependencies over a long period of time includes: constructing and training a temporal autoencoder, which includes a 1D-CNN layer, a max-pooling layer and a Bi-LSTM layer connected in sequence; using the trained temporal autoencoder to process each preprocessed load segment, extracting short-term features and variable coupling relationships of each load segment, and capturing sequential dependencies over a long period of time.
[0011] As a further defined implementation, the training process of the temporal autoencoder includes: taking all preprocessed load segments as input, using the minimum reconstruction mean square error as the loss function, and performing unsupervised pre-training with the Adam optimizer until the required number of training iterations is met.
[0012] As an alternative implementation, the process of reconstructing the original load sequence from the extracted feature vectors is implemented using a decoder. The decoder has a structure symmetrical to the encoder, and is connected in sequence with an upsampling layer, a deconvolution layer, and an LSTM layer. The goal is to reconstruct the original load segment sequence from the feature vectors.
[0013] As an alternative implementation, the process of calculating the morphological similarity between feature vectors extracted from each load segment includes: based on the low-dimensional feature vectors of each load segment... ; Perform the FastDTW algorithm calculation for each pair of feature vectors. and Multi-level coarse-grained processing is performed to generate low-resolution sequences; The initial curved path is found using standard DTW at the lowest resolution layer, and then gradually refined to a higher resolution layer to optimize distance calculation within the path neighborhood. Calculate the FastDTW distance between all sample pairs to form Symmetric distance matrix , of which elements The formula for calculating the cumulative distance path between load segments i and j, representing the morphological differences between the two sequences, is as follows:
[0014] in It is a regular path. It is a point-to-point distance. FastDTW reduces computational complexity by approximating the path with coarse-grained approximation.
[0015] As an alternative implementation, the process of clustering the load segment into K clusters includes: setting the range of values for the candidate cluster number K, performing preliminary clustering for each K value, and obtaining the cluster division; For each load segment sample i, calculate its profile coefficient. ; Intra-cluster dissimilarity Let the average FastDTW distance between sample i and all other samples in the same cluster be:
[0016] in It is the cluster to which sample i belongs. It is the cluster size; Intra-cluster dissimilarity Let i be the minimum average distance between sample i and all samples in other clusters.
[0017] Profile coefficient for:
[0018] The value of is in the range of [-1, 1], and the closer the value is to 1, the better the clustering effect; For each K, calculate the average of the silhouette coefficients for all samples:
[0019] Choose to The largest K is the optimal cluster number:
[0020] Based on certainty The K-Medoids algorithm is used to cluster and divide the load segments.
[0021] As an alternative implementation, the process of optimizing the parameters of the autoregressive integral moving average model using the artificial lemming algorithm and then using the optimized autoregressive integral moving average model to perform initial forecasting of the load sequence includes: Initialize the artificial lemming algorithm population with the fitness function that minimizes the AIC value of the autoregressive integral moving average model; In the migration behavior simulation, the distance between individuals is measured based on the FastDTW algorithm to calculate the similarity between individuals. Each individual migrates to a neighboring individual with higher fitness, and the migration probability is proportional to the fitness difference. Set up an elimination mechanism to eliminate the M individuals with the worst fitness in each generation; New parameter combinations are generated through crossover operations, and diversity is introduced through mutation operations; Perform iterative optimization until the convergence judgment condition or termination condition is met to obtain the optimal parameter combination. Use the optimal parameter combination obtained by optimization to construct an autoregressive integral moving average model. The historical load sequences within the clusters are input into the autoregressive integral moving average model for fitting, and the initial predicted values within the load segments are obtained to reflect the linear components and short-term dependencies of the sequences.
[0022] A short-term power load forecasting system for industrial parks includes: The segmentation and preprocessing module is configured to acquire load data from the industrial park, construct load curves, and segment and preprocess the load curves. The feature extraction module is configured to extract short-term features and variable coupling relationships of each load segment, and capture the sequential dependencies over a long time range, reconstructing the original load sequence from the extracted feature vectors. The clustering module is configured to calculate the morphological similarity between the feature vectors extracted from each load segment, and based on the calculated similarity, cluster the load segments into K clusters. The prediction module is configured to optimize the parameters of the autoregressive integral moving average model using the artificial lemming algorithm, use the optimized autoregressive integral moving average model to make an initial prediction of the load sequence, obtain the predicted value reflecting the linear component of the sequence, calculate the residual sequence between the predicted value and the actual load value, and use a gated cyclic unit to capture the nonlinear dynamics and long-term dependencies in the sequence to predict future residual values. The load forecast calculation module is configured to add the forecast value reflecting the linear component of the sequence to the future residual value to obtain the final load forecast value.
[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Effectively address the lack of historical data and improve the prediction capability of small samples: Through unsupervised pre-training and feature extraction by time-series autoencoders, deep time-series patterns can be effectively learned from small sample historical load data; combined with the "segmented clustering" strategy, sparse data is divided into more regular typical electricity consumption pattern clusters, so that reliable prediction can still be achieved in the park environment with limited data, reducing the model's dependence on large data volume.
[0024] (2) It has strong generalization and universality, and is applicable to various differentiated industrial parks: It does not rely on specific load composition or fixed electricity consumption patterns. Through data-driven cluster analysis, it automatically identifies a variety of typical electricity consumption patterns within the park and independently constructs the optimal prediction model for each pattern. This "cluster-based" framework enables it to adapt to the complex and variable load characteristics of different types of industrial parks (such as high-energy-consuming types, high-rise buildings, etc.) and has wide applicability.
[0025] (3) Achieving parameter self-optimization and synergistic improvement in accuracy: For each load cluster, the Artificial Lemming Algorithm (ALA) is used to automatically search for the optimal parameters of the ARIMA model, and the residuals are corrected by combining the GRU neural network. This hybrid model not only ensures accurate fitting of linear trends, but also effectively captures complex nonlinear fluctuations, realizing intelligent prediction process and high accuracy of prediction results.
[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0028] Figure 1 This is a flowchart of a method for predicting short-term power load in an industrial park, as exemplified by one such method. Figure 2 This is a flowchart of a Savitzky-Golay filtering smoothing process according to one embodiment; Figure 3 This is a flowchart of a feature extraction process based on a temporal autoencoder, according to one embodiment. Figure 4 This is a flowchart of a load segment clustering analysis based on FastDTW, as exemplified by one of the embodiments. Figure 5 This is a schematic diagram illustrating the construction process of an ALA+ARIMA+GRU hybrid prediction model in one embodiment. Detailed Implementation
[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0030] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0031] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. 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.
[0032] Where there is no conflict, the embodiments and features described in this application may be combined with each other.
[0033] Example 1 As described in the background section, most existing load forecasting methods are unable to effectively address the following three core challenges: (1) the complexity of the load curve shape caused by artificial means in response to peak-valley flat electricity pricing; (2) the weakening of the reference value of historical data due to rapid product updates; and (3) the lack of model training data due to incomplete data records.
[0034] To address the problems of complex load curve shapes caused by peak-valley-flat electricity pricing, reduced historical data reliability due to rapid product upgrades, and insufficient model training data due to incomplete data records, this embodiment provides a short-term power load forecasting method for industrial parks. Figure 1 As shown, it includes the following steps: Taking 22 industrial parks distributed throughout a province in a certain province from 2021 to 2024 as an example, the types of products processed and manufactured in each industrial park are shown in Table 1: Table 1
[0035] 1. Load curve segmentation and preprocessing 1.1. Time Period Division Based on the local standards for dividing peak, flat, and valley periods, a complete daily electricity load curve (one point every 1 minute) is divided into multiple independent load segments according to time points.
[0036] Based on the deep valley, low valley, flat, peak, and peak periods of industrial and commercial electricity prices, historical daily load data is segmented into load segments, and an "electricity price type" attribute is added to the segmented load curves. The segmentation types, combined with the province's electricity prices, are shown in Table 2. Table 2
[0037] 1.2. Data Smoothing Processing For each load segment, a Savitzky-Golay filter with a window length of 5 and a polynomial order of 2 is applied for smoothing to eliminate random noise interference while better preserving the local extrema and waveform characteristics of the load curve, such as... Figure 2 As shown.
[0038] 1.3. Data Alignment Processing Because the duration of peak, flat, and valley periods varies across different years and seasons, the number of data points for each load segment is inconsistent. To address this, the target length N is set as the number of points on the longest curve within this type of load segment (e.g., peak load segment). Then, linear interpolation is used to transform all curves into a unified sequence with N equally spaced time points.
[0039] 1.4. Normalization Process The minimum-maximum normalization method is used to normalize the load segment. The value of each curve is scaled to the [0, 1] interval, and the calculation formula is as follows:
[0040] in and These are the maximum and minimum values of the curve itself.
[0041] 2. Feature extraction based on temporal autoencoders 2.1. Model Building and Training The encoder is composed of a 1D-CNN layer (64 kernels, size 3, activation function 'ReLU'), a max pooling layer (pooling size 2), and a Bi-LSTM layer (32 hidden units), which outputs a fixed-length low-dimensional feature vector.
[0042] Decoder: Symmetrical to encoder, it is connected in sequence with an upsampling layer, a deconvolution layer, and an LSTM layer. Its goal is to reconstruct the original load segment sequence from the feature vector.
[0043] Training process: All preprocessed load segments are used as input, and the Adam optimizer is used for unsupervised pre-training with the loss function of minimizing the mean squared error of reconstruction (MSE). The training epochs are set to 100. After training, the decoder is discarded, and only the encoder is retained for feature extraction.
[0044] 2.2. Feature Vector Acquisition Each preprocessed load segment is input into the trained encoder to obtain its corresponding low-dimensional feature vector, which serves as the deep temporal feature representation of that load segment. Figure 3 As shown.
[0045] Step 1: The preprocessed and aligned load segment input data is as follows:
[0046] Where T is the unified time step (number of points).
[0047] Step 2: Calculation of a one-dimensional convolutional neural network (1D-CNN) Extracting local short-term features. Each convolutional kernel Slide the input sequence, compute its convolution value, and then add a bias. And through the activation function.
[0048] For the k-th convolutional kernel, the output at time step t for
[0049] After processing with 64 convolutional kernels, a new sequence with 64 channels is obtained:
[0050] Where L is the length of the sequence after convolution. .
[0051] Step 3: Max Pooling Calculation Reduce data dimensionality, enhance feature translation invariance, and reduce computational cost. For each channel, take the maximum value between two adjacent time steps. For the k-th channel, the output at pooling window j... for:
[0052] After calculation, the sequence length is further reduced to half of the original length, resulting in a new sequence:
[0053] Step 4: Bidirectional Long Short-Term Memory (Bi-LSTM) network calculation Capture contextual dependencies over a long period of time.
[0054] Forward LSTM: This method uses pooled sequences... ,in ,from arrive Process sequentially to obtain the final hidden state. .
[0055] Backward LSTM: Transforms the same sequence from arrive Reverse the process to obtain the final hidden state. .
[0056] The forward final hidden state and backward final hidden state Connect the features to obtain the final low-dimensional feature vector. .
[0057]
[0058] 3. Load segment cluster analysis based on FastDTW 3.1. Similarity Calculation Using the eigenvectors obtained in section 2.2, the FastDTW distance between any two load segments is calculated, and a distance matrix is constructed. The radius parameter of FastDTW is set to 5 to balance computational accuracy and efficiency.
[0059] Step 1: Extract the low-dimensional feature vectors of each load segment obtained in step 2.2. As input.
[0060] Step 2: FastDTW algorithm calculation. For each pair of feature vectors... and Multi-level coarse-grained processing is performed to generate low-resolution sequences.
[0061] The initial curved path is found using standard DTW at the lowest resolution layer. The process is then progressively refined to higher resolution layers, optimizing distance calculations within the path neighborhood (radius = 5).
[0062] Calculate the FastDTW distance between all sample pairs to form Symmetric distance matrix , of which elements This represents the morphological difference between load segments i and j (the smaller the distance, the higher the similarity). The formula for calculating the path that minimizes the cumulative distance between two sequences is:
[0063] in It is a regular path. It is a point-to-point distance. FastDTW reduces computational complexity by approximating the path with coarse-grained approximation.
[0064] 3.2. Determining the Optimal Number of Clusters The silhouette coefficient method was used to evaluate the clustering effect under different K values (K from 2 to 10), and the K value with the largest silhouette coefficient was selected as the optimal number of clusters.
[0065] Step 1: Set the range of K values: Set the range of candidate cluster number K to 2 to 10. For each K value, perform preliminary clustering to obtain cluster partitions.
[0066] Step 2: For each load segment sample i, calculate its profile coefficient. .
[0067] Intra-cluster dissimilarity : The average FastDTW distance between sample i and all other samples in the same cluster.
[0068]
[0069] in It is the cluster to which sample i belongs. It refers to the cluster size.
[0070] Intra-cluster dissimilarity : The minimum average distance between sample i and all samples in other clusters.
[0071]
[0072] Profile coefficient :
[0073] The value range is [-1, 1], and the closer the value is to 1, the better the clustering effect.
[0074] Step 3: Calculate the average silhouette coefficient. For each K, calculate the average silhouette coefficient of all samples:
[0075] Step 4: Choosing the optimal K value. Select the value that makes... The largest K is the optimal cluster number:
[0076] 3.3. Clustering Execution Based on certainty The K-Medoids algorithm (based on the PAM algorithm) is used to cluster the load segments. The center of each cluster is an actual load segment object, and each cluster represents a typical power consumption pattern.
[0077] Step 1: Randomly select K initial Medoids (cluster centers) .
[0078] Step 2: For each non-Medoid sample Assign to the nearest cluster containing the Medoid element, using the following formula:
[0079] in It is the cluster label of sample i. It is the FastDTW distance.
[0080] Cluster partitioning ,in .
[0081] Step 3: Calculate the current total cost TC, which represents the sum of intra-cluster dissimilarity:
[0082] Iterate through each Medoid Non-Medoid points Temporary use replace This yields a new Medoid set. Reassign all samples to the new Medoids and calculate the new total cost. .
[0083] Calculate the cost changes resulting from the exchange:
[0084] Choose the optimal swap, find the one that makes The swap pair with the largest negative value (i.e., the greatest cost reduction):
[0085] Iterative updates, if If the condition is met, then perform the swap, update the Medoids and cluster partitions; otherwise, the algorithm converges.
[0086] Step 4: The algorithm stops when the total cost change is less than the threshold in consecutive iterations or when no cost-reducing exchange can be found.
[0087] like Figure 4 As shown in Figure 4, construct a hybrid prediction model of ALA+ARIMA+GRU. For each cluster obtained in step 3, perform the following operations: 4.1. Optimize ARIMA parameters using ALA Step 1: Construct the ARIMA model and set the initial search range of the parameters (p, d, q) of the ARIMA model. The autoregression order p: [0,5], the difference order d: [0,2], and the moving average order q: [0,5].
[0088] ARIMA model:
[0089] Where B is the lag operator, i.e. ; It is an autoregressive polynomial; It is a moving average polynomial; It is a white noise sequence; d is the difference order, which makes the sequence stationary.
[0090] Step 2: Initialize the Artificial Lemming Algorithm (ALA) population to minimize the AIC value of the ARIMA model as a fitness function.
[0091] The fitness function is set to minimize the AIC (Akaike Information Criterion) value of the ARIMA model. AIC is used to balance model complexity and goodness of fit.
[0092] Where k is the total number of model parameters (i.e., p+q+1), L is the model likelihood function value, and ALA iteratively optimizes with the goal of minimizing AIC.
[0093] The fitness function is obtained as follows: (The lower the AIC value, the higher the fitness.) Step 3: Migration Behavior Simulation. Inter-individual distance measurement, calculating inter-individual similarity based on the FastDTW algorithm in section 3.1:
[0094] in, and Feature vectors corresponding to different parameter combinations The distance is approximated by multi-level coarse-grained reduction to approximate the classic DTW distance
[0095] The migration direction is determined, and each individual migrates to a neighboring individual with higher fitness. The migration probability is proportional to the fitness difference.
[0096] Step 4: Simulate competitive behavior. Set up an elimination mechanism, eliminating the M individuals with the worst fitness in each generation:
[0097] Where N represents the number of individuals maintained during the iteration process.
[0098] The elimination criteria are:
[0099] New individuals are generated by crossover operations to create new parameter combinations:
[0100] Introducing diversity through mutation operations:
[0101] in This represents a single entity, that is, a combination of ARIMA parameters. ; This represents the operation of combining the characteristics of two parent individuals to generate offspring, i.e., from... and Randomly select some parameters and combine them into a new parameter combination; This represents a mutation operation, applied to a single... The parameters are randomly perturbed to introduce diversity.
[0102] and express The father generation, For cross operation offspring, The new version modified by the mutation operation ; This represents the range of variation.
[0103] Step 5: Convergence Judgment and Termination Conditions. The process terminates when either of the following two conditions is met: Condition 1: The number of iterations reaches a preset value; Condition 2: The change in AIC is less than the threshold, which is set to 0.001.
[0104] Step 6: After the algorithm terminates, the optimal parameter combination is obtained:
[0105] 4.2. Linear prediction using the ARIMA model Step 1: Optimize the parameters obtained in step 4.1 Construct an ARIMA model.
[0106] Step 2: Input the historical load sequence within the cluster into the ARIMA model for fitting to obtain the initial predicted values within the load segment. It is used to reflect the linear components and short-term dependencies of a sequence.
[0107] 4.3. GRU Residual Correction Step 1: Residual Calculation. Calculate the residual sequence between the predicted and actual values of the ARIMA model on historical data. The residual formula is:
[0108] in, This represents the actual load value. These are ARIMA predicted values.
[0109] Step 2: GRU model training: Construct a two-layer GRU network (64 hidden units per layer), using the historical residual sequence as input and output target, and train the GRU model to learn the nonlinear patterns contained in the residuals.
[0110] Construct a two-layer GRU network, with 64 hidden units in each layer.
[0111] The input is a historical residual sequence. The output is the predicted residual. GRU captures long-term dependencies through gating mechanisms (reset gates and update gates).
[0112] The GRU model is trained with the objective of minimizing the mean squared error (MSE) of the prediction residuals.
[0113] Step 3: Residual Prediction: Using the trained GRU model, predict the residual values for the corresponding future step size. .
[0114] 4.4. Result Synthesis: The corresponding points of the linear prediction results of the ARIMA model and the nonlinear residual prediction results of the GRU model are added together to obtain the final load prediction value under this electricity consumption pattern cluster.
[0115]
[0116] 5. Model Performance Evaluation and Comparison The dataset is divided into a training set and a test set in chronological order (80% for training and 20% for testing). On the test set, the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the hybrid prediction model of this invention are calculated.
[0117] Example 2 A short-term power load forecasting system for industrial parks includes: The segmentation and preprocessing module is configured to acquire load data from the industrial park, construct load curves, and segment and preprocess the load curves. The feature extraction module is configured to extract short-term features and variable coupling relationships of each load segment, and capture the sequential dependencies over a long time range, reconstructing the original load sequence from the extracted feature vectors. The clustering module is configured to calculate the morphological similarity between the feature vectors extracted from each load segment, and based on the calculated similarity, cluster the load segments into K clusters. The prediction module is configured to optimize the parameters of the autoregressive integral moving average model using the artificial lemming algorithm, use the optimized autoregressive integral moving average model to make an initial prediction of the load sequence, obtain the predicted value reflecting the linear component of the sequence, calculate the residual sequence between the predicted value and the actual load value, and use a gated cyclic unit to capture the nonlinear dynamics and long-term dependencies in the sequence to predict future residual values. The load forecast calculation module is configured to add the forecast value reflecting the linear component of the sequence to the future residual value to obtain the final load forecast value.
[0118] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).
[0119] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0120] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0121] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for short-term power load forecasting in industrial parks, characterized in that, Includes the following steps: Obtain load data from the industrial park, construct load curves, and segment and preprocess the load curves. Extract short-term features and variable coupling relationships for each load segment, and capture the sequential dependencies over a long time period. Reconstruct the original load sequence from the extracted feature vectors. Based on the feature vectors extracted from each load segment, the morphological similarity between each feature vector is calculated. Based on the calculated similarity, the load segments are clustered into K clusters. The parameters of the autoregressive integral moving average model are optimized using the artificial lemming algorithm. The optimized autoregressive integral moving average model is used to make an initial prediction of the load sequence to obtain the predicted value reflecting the linear component of the sequence. The residual sequence between the predicted value and the actual load value is calculated. The nonlinear dynamics and long-term dependencies in the sequence are captured by the gated cyclic unit to predict the future residual value. The predicted value, which reflects the linear component of the sequence, is added to the future residual value to obtain the final load forecast.
2. The method for short-term power load forecasting in industrial parks as described in claim 1, characterized in that, The process of segmenting the load curve includes dividing the complete daily electricity load curve into multiple independent load segments according to the peak, flat and valley time period division standards issued by the region to which the industrial park is located.
3. The method for short-term power load forecasting in industrial parks as described in claim 1, characterized in that, The process of preprocessing the load curve includes: Savitzky-Golay filters are used to denoise each load segment while preserving local morphological features. Linear interpolation is used to unify load segments of different lengths to the number of points on the longest curve in the set, thus achieving data alignment. Min-Max normalization is used to scale the load values to the [0,1] interval for normalization, eliminating the influence of dimensions.
4. The method for short-term power load forecasting in industrial parks as described in claim 1, characterized in that, The process of extracting short-term features and variable coupling relationships of each load segment and capturing sequential dependencies over a long period of time includes: constructing and training a temporal autoencoder, which includes a 1D-CNN layer, a max pooling layer and a Bi-LSTM layer connected in sequence; using the trained temporal autoencoder to process each preprocessed load segment, extracting short-term features and variable coupling relationships of each load segment, and capturing sequential dependencies over a long period of time.
5. The method for short-term power load forecasting in industrial parks as described in claim 4, characterized in that, The training process of a time-series autoencoder includes: taking all preprocessed load segments as input, using the minimum reconstruction mean square error as the loss function, and performing unsupervised pre-training with the Adam optimizer until the required number of training iterations is met.
6. The method for short-term power load forecasting in industrial parks as described in claim 1, characterized in that, The process of reconstructing the original load sequence from the extracted feature vectors is implemented using a decoder. The decoder has a structure symmetrical to the encoder, and is connected in sequence with an upsampling layer, a deconvolution layer, and an LSTM layer. The goal is to reconstruct the original load segment sequence from the feature vectors.
7. The method for short-term power load forecasting in industrial parks as described in claim 1, characterized in that, The process of calculating the morphological similarity between feature vectors extracted from each load segment includes: based on the low-dimensional feature vectors of each load segment... ; Perform the FastDTW algorithm calculation for each pair of feature vectors. and Multi-level coarse-grained processing is performed to generate low-resolution sequences; The initial curved path is found using standard DTW at the lowest resolution layer, and then gradually refined to a higher resolution layer to optimize distance calculation within the path neighborhood. Calculate the FastDTW distance between all sample pairs to form Symmetric distance matrix , of which elements The formula for calculating the cumulative distance path between load segments i and j, representing the morphological differences between the two sequences, is as follows: in It is a regular path. It is a point-to-point distance. FastDTW approximates this path with coarse-grained approximation to reduce computational complexity.
8. The method for short-term power load forecasting in industrial parks as described in claim 1, characterized in that, The process of clustering the load segment into K clusters includes: setting the range of values for the candidate cluster number K, performing preliminary clustering for each K value, and obtaining the cluster division; For each load segment sample i, calculate its profile coefficient. ; Intra-cluster dissimilarity Let the average FastDTW distance between sample i and all other samples in the same cluster be: in It is the cluster to which sample i belongs. It is the cluster size; Intra-cluster dissimilarity Let i be the minimum average distance between sample i and all samples in other clusters. Profile coefficient for: The value of is in the range of [-1, 1], and the closer the value is to 1, the better the clustering effect; For each K, calculate the average of the silhouette coefficients for all samples: Choose to The largest K is the optimal cluster number: Based on certainty The K-Medoids algorithm is used to cluster and divide the load segments.
9. The method for short-term power load forecasting in industrial parks as described in claim 1, characterized in that, The process of optimizing the parameters of the autoregressive integral moving average model using the artificial lemming algorithm and then using the optimized autoregressive integral moving average model to make initial predictions on the load series includes: Initialize the artificial lemming algorithm population with the fitness function that minimizes the AIC value of the autoregressive integral moving average model; In the migration behavior simulation, the distance between individuals is measured based on the FastDTW algorithm to calculate the similarity between individuals. Each individual migrates to a neighboring individual with higher fitness, and the migration probability is proportional to the fitness difference. Set up an elimination mechanism to eliminate the M individuals with the worst fitness in each generation; New parameter combinations are generated through crossover operations, and diversity is introduced through mutation operations; Perform iterative optimization until the convergence judgment condition or termination condition is met to obtain the optimal parameter combination. Use the optimal parameter combination obtained by optimization to construct an autoregressive integral moving average model. The historical load sequences within the clusters are input into the autoregressive integral moving average model for fitting, and the initial predicted values within the load segments are obtained to reflect the linear components and short-term dependencies of the sequences.
10. A short-term power load forecasting system for industrial parks, characterized in that, include: The segmentation and preprocessing module is configured to acquire load data from the industrial park, construct load curves, and segment and preprocess the load curves. The feature extraction module is configured to extract short-term features and variable coupling relationships of each load segment, and capture the sequential dependencies over a long time range, reconstructing the original load sequence from the extracted feature vectors. The clustering module is configured to calculate the morphological similarity between the feature vectors extracted from each load segment, and based on the calculated similarity, cluster the load segments into K clusters. The prediction module is configured to optimize the parameters of the autoregressive integral moving average model using the artificial lemming algorithm, use the optimized autoregressive integral moving average model to make an initial prediction of the load sequence, obtain the predicted value reflecting the linear component of the sequence, calculate the residual sequence between the predicted value and the actual load value, and use a gated cyclic unit to capture the nonlinear dynamics and long-term dependencies in the sequence to predict future residual values. The load forecast calculation module is configured to add the forecast value reflecting the linear component of the sequence to the future residual value to obtain the final load forecast value.