Time sequence prediction method and system based on multi-scale feature reconstruction and multi-expert fusion

Abstract

The invention relates to the technical field of machine learning, and particularly provides a time sequence prediction method and system based on multi-scale feature reconstruction and multi-expert fusion, and the method comprises the steps: obtaining and preprocessing multi-source time sequence data, and constructing a structured feature matrix; performing multi-scale feature reconstruction on the sequence in the matrix by using three groups of one-dimensional convolutional networks with different receptive fields to obtain reconstructed feature sample sets with local, medium and global scales; performing time sequence slicing on the sample set, and calculating statistical characteristics in each slice; for the sample slices of each scale, respectively obtaining three prediction components of basic linear extrapolation, covariant linear correction and nonlinear residual through parallel branches; based on the statistical characteristics, dynamically fusing the three components through a gating network to obtain branch prediction results of all scales; and finally, fusing all branch prediction results through a full-connection network, and outputting a final prediction value. According to the method, the accuracy, robustness and practicability of time sequence prediction are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to a time series prediction method and system based on multi-scale feature reconstruction and multi-expert fusion. Background Technology

[0002] Multi-input single-parameter time-series forecasting technology has significant application value in fields such as power load forecasting and power generation forecasting. However, existing methods suffer from significant shortcomings when dealing with complex time-series data: traditional statistical models, such as vector autoregression, rely on linear assumptions and struggle to characterize nonlinear interactions; deep learning models, such as recurrent neural networks and their variants, while capable of handling sequence dependencies, suffer from low training efficiency and vanishing gradients; the Transformer architecture based on self-attention has made progress in long-range dependency modeling, but when handling multivariate inputs, it typically mixes different variables along the feature dimension, making it difficult to explicitly extract physically meaningful multi-scale features, and lacks adaptive modeling mechanisms for the dynamic relationships between the target variable and different covariates. Furthermore, existing methods generally suffer from common drawbacks such as single-granularity feature extraction, static fusion mechanisms, and low residual learning efficiency, limiting their performance in multi-scale adaptive feature extraction, dynamic fusion of multi-branch results, and efficient nonlinear residual learning. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention provides a time series prediction method and system based on multi-scale feature reconstruction and multi-expert fusion to solve the above-mentioned technical problems.

[0004] In a first aspect, the present invention provides a time series prediction method based on multi-scale feature reconstruction and multi-expert fusion, comprising: Acquire multi-source time-series data containing target sequence and covariate sequence, preprocess the multi-source time-series data and construct a structured feature matrix containing multi-dimensional time features; Using three sets of one-dimensional convolutional networks with different receptive fields, multi-scale feature reconstruction is performed on the sequences in the feature matrix to obtain reconstructed feature sample sets of local, intermediate and global scales; The reconstructed feature sample set is divided into time-series slices, and the statistical features within each slice are calculated; For each scale of samples after time-series slicing, multiple prediction components are obtained through parallel branches, including: basic linear extrapolation components based on historical target sequences, linear correction components based on future covariate sequences, and nonlinear residual components based on the splicing of historical and future sequences. Based on the statistical characteristics, multiple prediction components are dynamically fused through a gating network to obtain branch prediction results at each scale; all branch results are then fused through a fully connected network to output the final prediction value.

[0005] In one optional implementation, the target sequence is an electricity load sequence or an electricity consumption sequence; the covariate sequence includes meteorological data and time characteristic data.

[0006] In an optional implementation, preprocessing the multi-source time-series data and constructing a structured feature matrix containing multi-dimensional temporal features includes: The multi-source time-series data is divided into training set, validation set and test set according to time order and a predetermined ratio; Encode the timestamps of the training set to construct a multidimensional time feature containing date and period attributes; The target sequence, covariate sequence, and multidimensional time features are concatenated to generate a structured feature matrix.

[0007] In an optional implementation, preprocessing the multi-source time-series data and constructing a structured feature matrix containing multi-dimensional temporal features includes: Time series alignment is performed based on the sampling characteristics of each data source, and training, validation and test sets are adaptively divided according to the volatility index of time series data. Timestamp information is input into a temporal feature encoding network, which automatically extracts and outputs multidimensional temporal features that fuse periodicity and event-related characteristics. The temporal feature encoding network includes a temporal autoencoder architecture. The encoder part of the temporal autoencoder includes an attention mechanism for capturing global dependencies and a convolutional module for capturing local periodic patterns. A feature relationship graph is constructed with the target sequence and covariate sequence as nodes. The node information is aggregated through a graph neural network to generate a feature matrix that encodes the correlation between variables, which serves as the structured feature matrix. The feature relationship graph is a fully connected graph, and its edge weights are initialized based on the statistical correlation or mutual information between variables.

[0008] In an optional implementation, three sets of one-dimensional convolutional networks with different receptive fields are used to reconstruct the sequences in the feature matrix at multiple scales, resulting in reconstructed feature sample sets at three scales: local, intermediate, and global, including: A three-layer, depth-increasing one-dimensional convolutional network is used, corresponding to the local, intermediate, and global scales, respectively. The first group of convolutional networks consists of single-layer convolutions and is used to extract detailed features at the local scale. The second set of convolutional networks consists of two cascaded convolutional layers and is used to extract the medium-scale periodic features. The third set of convolutional networks consists of three cascaded convolutional layers and is used to extract the global-scale trend features.

[0009] In an optional implementation, the reconstructed feature sample set is time-series sliced, and statistical features within each slice are calculated, including: Using a preset fixed length as the history window, the reconstructed feature sample set is sliced ​​by sliding along the time axis with a step size of 1. For each slice, calculate statistics for the target sequence and at least one key covariate within its historical window, including at least one of the mean, standard deviation, and maximum value.

[0010] In an optional implementation, for each scale of samples after time-series slicing, multiple prediction components are obtained through parallel branches, including: Input the historical target sequence of each scale sample into a basic linear projection layer to obtain the basic linear extrapolation component; Input the future covariate sequence of each scale sample into a covariate linear projection layer to obtain the covariate linear correction component; The historical target sequence, historical covariate sequence, and future covariate sequence of each scale sample are concatenated into a conditional sequence, which is then input into a conditional autoencoder based on a multilayer perceptron to obtain nonlinear residual components.

[0011] In an optional implementation, based on the statistical features, multiple prediction components are dynamically fused through a gating network to obtain branch prediction results at each scale, including: The statistical features are input into a gating network, and the gating network outputs dynamic weights corresponding to the basic linear extrapolation components, covariate linear correction components, and nonlinear residual components, respectively. After normalizing the dynamic weights, the results of the three prediction components are weighted and summed to obtain the branch prediction results at the corresponding scale.

[0012] In an optional implementation, the gated network includes a multilayer perceptron, which maps the statistical features of the input to the initial weights of each predicted component through a nonlinear transformation, and then normalizes them using a Softmax function to output the dynamic weights with a sum of 1.

[0013] Secondly, the present invention provides a time series prediction system based on multi-scale feature reconstruction and multi-expert fusion, comprising: The data acquisition module is used to acquire multi-source time-series data containing target sequences and covariate sequences, preprocess the multi-source time-series data, and construct a structured feature matrix containing multi-dimensional time features. The feature reconstruction module is used to perform multi-scale feature reconstruction on the sequence in the feature matrix using three sets of one-dimensional convolutional networks with different receptive fields, so as to obtain three sets of reconstructed feature sample sets at local, medium and global scales. The slicing module is used to perform time-series slicing on the reconstructed feature sample set and calculate the statistical features within each slice; The branch prediction module is used to obtain multiple prediction components for samples after time-series slicing at each scale through parallel branches, including: basic linear extrapolation components based on historical target sequences, linear correction components based on future covariate sequences, and nonlinear residual components based on the splicing of historical and future sequences. The branch fusion module is used to dynamically fuse multiple prediction components through a gating network based on the statistical features to obtain branch prediction results at each scale; and to fuse all branch results through a fully connected network to output the final prediction value.

[0014] Thirdly, a device is provided, comprising: The memory is used to store time series prediction programs based on multi-scale feature reconstruction and multi-expert fusion; The processor is configured to implement the steps of the time series prediction method based on multi-scale feature reconstruction and multi-expert fusion as provided in the first aspect when executing the time series prediction program based on multi-scale feature reconstruction and multi-expert fusion.

[0015] Fourthly, a computer-readable storage medium is provided, on which a time series prediction program based on multi-scale feature reconstruction and multi-expert fusion is stored. When the time series prediction program based on multi-scale feature reconstruction and multi-expert fusion is executed by a processor, it implements the steps of the time series prediction method based on multi-scale feature reconstruction and multi-expert fusion provided in the first aspect.

[0016] The beneficial effects of this invention are as follows: The time series prediction method and system based on multi-scale feature reconstruction and multi-expert fusion provided by this invention explicitly separates and extracts the local, periodic, and trend components of time series data through a multi-scale feature reconstruction mechanism, realizing a complete characterization of cross-scale dynamic characteristics; the introduction of a gating network based on data statistical features can adaptively adjust the fusion weights of each prediction component according to the real-time characteristics of the input sequence, significantly improving the adaptability and prediction accuracy of the model in complex and variable scenarios; at the same time, a well-structured MLP autoencoder is used to specifically learn nonlinear residuals, achieving a good balance between computational efficiency and model simplicity while ensuring strong fitting ability, thereby improving the overall accuracy, robustness, and practicality of time series prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0019] Figure 2 This is another illustrative flowchart of a method according to an embodiment of the present invention.

[0020] Figure 3 This is a flowchart of a multi-expert prediction branch MLP-based autoencoder method according to an embodiment of the present invention.

[0021] Figure 4 This is a schematic block diagram of a system according to an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

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

[0024] Unless otherwise defined, 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. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0025] The time series prediction method based on multi-scale feature reconstruction and multi-expert fusion provided in this embodiment of the invention is executed by a computer device, and correspondingly, the time series prediction system based on multi-scale feature reconstruction and multi-expert fusion runs in the computer device.

[0026] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be a time series prediction system based on multi-scale feature reconstruction and multi-expert fusion. Depending on different requirements, the order of the steps in this flowchart can be changed, and some can be omitted.

[0027] like Figure 1 As shown, the method includes: S1. Obtain multi-source time-series data containing the target sequence and covariate sequences, preprocess the multi-source time-series data, and construct a structured feature matrix containing multi-dimensional time features; S2. Using three sets of one-dimensional convolutional networks with different receptive fields, multi-scale feature reconstruction is performed on the sequences in the feature matrix to obtain reconstructed feature sample sets of local, intermediate and global scales; S3. Perform time-series slicing on the reconstructed feature sample set and calculate the statistical features within each slice; S4. For samples after time-series slicing at each scale, multiple prediction components are obtained through parallel branches, including: basic linear extrapolation components based on historical target sequences, linear correction components based on future covariate sequences, and nonlinear residual components based on the splicing of historical and future sequences. S5. Based on the statistical characteristics, multiple prediction components are dynamically fused through a gating network to obtain branch prediction results at each scale; all branch results are fused through a fully connected network to output the final prediction value.

[0028] Please refer to Figure 2 The detailed steps of this method are explained below.

[0029] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0030] Based on the target power load forecasting scenario, the following relevant time range and measurement point set were determined. The data used is a processed, anonymized power load dataset from a certain region. The time range is selected as the 18-month period from January 1, 2024 to July 1, 2025, with a data collection interval of 15 minutes, and 96 data points collected per day.

[0031] The power load forecasting model selected five operational data points (including the date column). In addition to the target data point load, there are four meteorological data points. The specific details of the data points are shown in Table 1 below: Table 1 Measurement Points for Power Load Forecasting Business Model

[0032] Obtain the complete historical dataset H according to the above requirements. The dataset size is 52512 rows and 6 columns. The historical dataset H is divided into a training set TR, a validation set V, and a test set TS by default in a ratio of 0.7:0.2:0.1. All measurement points (except dates) in the training set TR are standardized according to the following formula:

[0033] in It is the mean of each measuring point. It is the standard deviation of each measuring point.

[0034] Step 1.2: Construction and scaling of temporal features The timestamp information of the training set TR is deep encoded to extract and construct 10 time features, as shown in Table 2. This process corresponds to the "time dimension expansion" operation in the figure.

[0035] Table 2. Time Characteristics of Dimensional Expansion

[0036] Considering the inconsistency in the units of different time features, feature scaling is required. All time feature values ​​must be scaled to [-0.5, 0.5] according to the scaling formula. The scaling formula is:

[0037] Where, normal represents the scaling result; t represents the time feature; baseValue is the baseline value, which defaults to 1; and rangeValue is the numerical fluctuation value.

[0038] Step 1.3: Generation of Structured Feature Matrix The processed target sequence (load feature), covariate sequence (meteorological feature N), and constructed multidimensional time feature M are concatenated column-wise to form the final structured multivariate time series matrix (FT). This matrix has dimensions [T, 1 + M + N], where T is the total number of rows in the training set TR (36758), 1 represents the target variable, M=10 represents the time feature dimension, and N=4 represents other covariate dimensions such as meteorological features. This FT matrix (with a shape and size of [36758, 15]) is the input data flowing to the "multi-scale feature reconstruction" module in the diagram.

[0039] First, data preprocessing and generation of structured feature matrices are performed. After obtaining the aligned original dataset, it is strictly divided into training, validation, and test sets in a 7:2:1 ratio according to time order to ensure the rigor of model evaluation. Next, the timestamp information of the training set is deeply encoded to construct multi-dimensional time features. The constructed features include: year, month, number of days in the month, day of the week (Monday is 0), hour, minute, whether it is a weekend, whether it is a holiday, season (value 1 to 4), and number of days in the year (Day of Year), totaling 10 features.

[0040] Finally, the standardized training set target load sequence, 4-dimensional meteorological covariate sequence, and the aforementioned 10-dimensional time features are concatenated along the feature dimensions to generate the final structured feature matrix. Assuming the time step of the training set is T, the dimensions of this matrix are [T, 15] (1-dimensional load + 4-dimensional meteorology + 10-dimensional time), and this matrix will serve as the input to the subsequent multi-scale feature reconstruction module.

[0041] To enable those skilled in the art to more fully understand the present invention, a detailed description is provided below in conjunction with another preferred embodiment. This embodiment focuses on demonstrating a more advanced adaptive and automated technical approach in the data preprocessing and feature construction stages.

[0042] First, dynamic time-series alignment and adaptive segmentation are performed. For multi-source data from different acquisition devices with non-uniform sampling frequencies (such as meteorological equipment sampling at the second level and load equipment sampling at the minute level), a dynamic time warping algorithm is used for sequence alignment to eliminate small offsets on the time axis. Next, the sliding window variance of the target sequence across all time periods is calculated as a volatility index. Based on this index, periods with significantly higher volatility than the average level are allocated a higher proportion to the training set (e.g., increased to 80%) to ensure the model fully learns complex patterns, while relatively stable periods are divided according to the baseline proportion, thus achieving adaptive partitioning of the dataset.

[0043] Secondly, deep encoding of temporal features is performed. The uniformly aligned timestamp sequence is input into a dedicated temporal feature encoding network. This network adopts a temporal autoencoder architecture, whose encoder part consists of two parallel branches: the first branch uses a multi-head self-attention mechanism to capture global dependencies caused by events such as dates and holidays; the second branch uses a set of dilated convolutional layers with different dilation coefficients, specifically for automatically detecting and extracting local periodic patterns such as daily and weekly cycles. The outputs of the two branches are fused along the feature dimension, and then compressed through a fully connected layer, ultimately automatically outputting a set of low-dimensional, dense, multi-dimensional temporal features that fuse periodicity and events, replacing traditional manual feature engineering.

[0044] Finally, cross-variable correlation feature fusion is implemented. The target load sequence, the sequences of each meteorological covariate, and the multidimensional time-series features obtained in the previous step are collectively regarded as feature nodes to construct a fully connected feature graph. The initial weight of each edge in the graph is determined by calculating the absolute value of the Pearson correlation coefficient or the mutual information value of the corresponding two variables on the entire training set. Subsequently, this graph is fed into a two-layer graph attention network (GNN), which learns and enhances the deep nonlinear correlations between variables through message passing and attention aggregation mechanisms. The output feature vectors of each node in the last layer of the GNN constitute the final structured feature matrix that encodes the correlations between variables.

[0045] In one embodiment of the present invention, based on step S2, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0046] This step corresponds to Figure 2 The core of the multi-scale feature reconstruction part on the left side lies in generating trend feature samples with different resolutions.

[0047] For the target variable (load) in the multivariate time series matrix FT and meteorological covariates Five sequences are each subjected to convolutional smoothing to extract trends at different scales. Definition A set of one-dimensional convolutional kernels, with the number of layers K (receptive field) increasing in each group, are set to... 3: Group 1 (Local Scale, L): Uses a 1-layer convolutional network with a kernel size of [missing value]. The pooling kernel size is It aims to capture short-term detailed features of the sequence.

[0048] Group 2 (Medium Scale, M): Uses a 2-layer convolutional network, with each layer having a kernel size of [size missing]. The pooling kernel size is The aim is to smooth sequence noise and extract sequence periodic features.

[0049] Group 3 (Global Scale, G): Uses a 3-layer convolutional network, with each kernel size being [size missing]. =5, pooling kernel size is It aims to capture the long-term trend of a sequence.

[0050] The convolution operation on sequence data of the target variable can be formally represented as: (3) in Indicates the first Group convolutional network functions, Given the input target variable sequence, For the reconstructed first The target sequence is sized, and T is the length of the target sequence.

[0051] Similarly, for each meteorological covariate Perform the same convolution operation to obtain Since convolution can shorten the sequence length, it needs to be restored to its original length using linear interpolation. The reconstructed target sequence, meteorological sequence, and original temporal feature covariates will be compared. By splicing, we obtain the first... Complete historical input samples at scale Where L represents the local scale, M represents the intermediate scale, and G represents the global scale.

[0052] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0053] Step 3.1: Preparation of the sliced ​​sample set The following uses historical sample data at a local scale. This example illustrates the process of obtaining a historical slice dataset containing a target variable (such as load) and covariates.

[0054] set up The total data duration is T, and the time interval is... (e.g., 15 minutes). Each time point The data is a vector containing one target value. The data consists of M time-related covariates (e.g., year, month, day, hour) and N meteorological covariates (e.g., temperature, humidity), meaning the data points have a dimension of M. Define the model input length S, prediction length P, and slice length used for training. With step size Sliding sampling is performed on the time axis to obtain a series of sample slices. For the first... 1 sample, whose input is from time The beginning of the continuous All data at each point in time are denoted as The corresponding prediction target is from Continuous from the start of time The target value at each time point is denoted as... For simplicity, the sample index will be omitted below. Finally, a local-scale slice sample set can be obtained. The total number of slices is T-L+1.

[0055] in The input section contains historical data in three parts: the target variable, time features, and meteorological covariates. , The output section also includes three parts of the target prediction data: the target variable, time characteristics, and meteorological covariates. .

[0056] Similarly, performing the same slicing operation on both the intermediate and global scales yields an intermediate-scale slice sample set. and global scale slice sample set .

[0057] Step 3.2: Calculation of statistical characteristics The following uses historical sample data at a local scale. Let's take an example to illustrate the calculation process of statistical characteristics.

[0058] Historical samples at local scale Input section Calculate the statistics of the target variable and key meteorological variables (such as temperature) within its internal window (e.g., the first S points). For example, calculate the average ( ), standard deviation ) and maximum value ( For a local scale L, the statistical eigenvectors can be obtained. : This vector will be used as input to the subsequent gating network.

[0059] in, These six values ​​are the mean, standard deviation, and maximum value of the target variable, and the mean, standard deviation, and maximum value of the key covariates.

[0060] Similarly, by performing the same statistical feature calculations on both the intermediate and global scales, we can obtain the intermediate-scale statistical feature vector. and global scale statistical eigenvectors The three feature vectors will serve as input data for the subsequent gating network.

[0061] In one embodiment of the present invention, based on step S4, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0062] This step corresponds to Figure 2 The middle multi-component / expert prediction section contains three parallel branches, each with the same structure. The following explanation will take the local scale L as an example.

[0063] Step 4.1: Basic Linear Extrapolation The input part of the target slice sequence at local scale L Projecting through a linear layer to directly predict the future The numerical value of the point.

[0064]

[0065] in, Indicates will Flattening a 1D vector This represents a linear mapping function that linearly maps an S-dimensional vector to a P-dimensional vector. This represents the P-dimensional vector output by the linear mapping function.

[0066] Step 4.2: Covariate linear correction Future at local scale L All covariates of a point Through another linear layer, the influence of covariates on the target variable is learned to predict the future value of the target point P.

[0067]

[0068] in, It is a linear transformation function applied point-in-time, transforming the input dimension into a linear transformation function. The vector is mapped to an output vector of dimension 1. This represents the P-dimensional vector output by the linear mapping function.

[0069] Step 4.3: Residual Nonlinearity Correction This part uses an MLP-based autoencoder (structure see...). Figure 2 Learning nonlinear residuals. First, construct the conditional input:

[0070] in, These are placeholders (such as zero vectors) reserved for historical target parts in future periods. This indicates concatenation along the time dimension, resulting in a concatenated sequence with a total length of [value missing]. . These refer to the input data of the target variable, the input data of the time feature, the input data of the meteorological variable, the future data of the time feature, and the future data of the meteorological variable, respectively.

[0071] Then, Input MLP-based autoencoder : .

[0072] Self-encoder The specific structure is as follows Figure 3 As shown, it includes an encoder and a decoder, both of which are composed of multilayer perceptrons (MLPs). The future of MLP-based autoencoder output Point prediction data. Both the encoder and decoder consist of two concatenated layers. The encoder output, after dimensionality transformation, is concatenated with the future covariate matrix along the feature dimension and input to the decoder for feature decoding. Both the encoding and decoding layers employ a feedforward neural network architecture, with its core consisting of a residual calculation unit and a linear mapping module. Specifically, the residual calculation unit is implemented through a cascaded operation of "linear transformation – activation function – linear transformation," and its output is then concatenated with the result of the linear mapping module, and after layer normalization, it is output to the next layer.

[0073] In one embodiment of the present invention, based on step S5, a possible embodiment will be given below, and its specific implementation will be described in a non-limiting manner.

[0074] Step 5.1: Gating Weight Generation and Fusion Local scale statistical eigenvectors Input a gated network (usually an MLP containing several fully connected layers and activation functions) and output normalized weights through the Softmax function.

[0075]

[0076] in, , These are the weighting coefficients for basic linear extrapolation, covariate linear correction, and residual nonlinear correction at the local scale, respectively. MLP is the designed multilayer perceptron network.

[0077] Finally, by weighting and summing the results of the three components, the prediction results at the local scale can be obtained:

[0078] in, These are the prediction results for future point P under local scale basic linear extrapolation, covariate linear correction, and residual nonlinear correction, respectively.

[0079] Similarly, the prediction component calculation is performed on both the intermediate scale M and the global scale G to obtain the corresponding prediction results. All of them The vector result.

[0080] Step 5.2: Multi-scale result fusion stage This step corresponds to Figure 2 The multi-scale fusion section on the right side of the middle section is used to fuse the prediction results of the three branches.

[0081] Multi-scale prediction results are stitched together. The prediction results of the parallel branches are concatenated column-wise along the feature dimension.

[0082]

[0083] in, , This indicates a feature dimension column concatenation operation. These are the prediction results at the local, intermediate, and global scales, respectively. This is the concatenated prediction result matrix.

[0084] Nonlinear fusion of prediction results: The spliced ​​tensor Remodeling Two-dimensional matrix ( (for batch size), and then through a final fully connected fusion network. Nonlinear fusion is performed to map the K-dimensional features back to the target dimension 1.

[0085]

[0086] in, This represents a multilayer perceptron network function that performs a nonlinear mapping from a K-dimensional input to a 1-dimensional output. This represents a dimensional transformation operation. The final output is the model's prediction of the future. A sequence of predicted target values ​​(e.g., load) at various time points.

[0087] Model parameter optimization based on the validation set: To obtain a model with strong generalization ability and prevent overfitting, continuous performance tuning should be performed using a validation set V, which accounts for 20% of the historical data.

[0088] Iterative Validation and Early Stopping Mechanism: The total number of training epochs for the training set TR is set to 10. After all iterations of training on the training dataset TR are completed in each epoch, the current model is used to perform forward propagation on the validation set V, and the root mean square error (RMSE) loss function value on the validation set V is calculated. This process is used to monitor the model's performance on unseen data. This invention introduces an early stopping mechanism: if the loss on the validation set V no longer decreases or even begins to increase within three consecutive iterations, the model is determined to be overfitted, training is terminated early, and the model parameters are rolled back to the snapshot with the best performance on the validation set V.

[0089] In some embodiments, the time series prediction system based on multi-scale feature reconstruction and multi-expert fusion may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the time series prediction system based on multi-scale feature reconstruction and multi-expert fusion may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) Functionality of time series prediction based on multi-scale feature reconstruction and multi-expert fusion.

[0090] In this embodiment, the time series prediction system based on multi-scale feature reconstruction and multi-expert fusion can be divided into multiple functional modules according to its functions, such as... Figure 4 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0091] The data acquisition module is used to acquire multi-source time-series data containing target sequences and covariate sequences, preprocess the multi-source time-series data, and construct a structured feature matrix containing multi-dimensional time features. The feature reconstruction module is used to perform multi-scale feature reconstruction on the sequence in the feature matrix using three sets of one-dimensional convolutional networks with different receptive fields, so as to obtain three sets of reconstructed feature sample sets at local, medium and global scales. The slicing module is used to perform time-series slicing on the reconstructed feature sample set and calculate the statistical features within each slice; The branch prediction module is used to obtain multiple prediction components for samples after time-series slicing at each scale through parallel branches, including: basic linear extrapolation components based on historical target sequences, linear correction components based on future covariate sequences, and nonlinear residual components based on the splicing of historical and future sequences. The branch fusion module is used to dynamically fuse multiple prediction components through a gating network based on the statistical features to obtain branch prediction results at each scale; and to fuse all branch results through a fully connected network to output the final prediction value.

[0092] Figure 5 The time series prediction method based on multi-scale feature reconstruction and multi-expert fusion provided in the embodiments of this application can be applied to devices. Those skilled in the art will understand that the device structures involved in the embodiments of this invention do not constitute a limitation on the device. A device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. In the embodiments of this invention, the device includes, but is not limited to, laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of this application described and / or claimed herein.

[0093] The device 500 may include a processor 510, a memory 520, and a communication unit 530. These components communicate via one or more buses. Those skilled in the art will understand that the server structure shown in the figure does not constitute a limitation of the present invention. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0094] The memory 520 can be used to store execution instructions of the processor 510. The memory 520 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. When the execution instructions in the memory 520 are executed by the processor 510, the device 500 is able to perform some or all of the steps in the above method embodiments.

[0095] The processor 510 serves as the control center of the storage device, connecting various parts of the electronic device via various interfaces and lines. It executes software programs and / or modules stored in the memory 520, and calls data stored in the memory to perform various functions of the electronic device and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 510 may consist only of a central processing unit (CPU). In this embodiment of the invention, the CPU may have a single processing core or include multiple processing cores.

[0096] The communication unit 530 is used to establish a communication channel, enabling the storage device to communicate with other devices. It can receive user data sent by other devices or send user data to other devices.

[0097] The present invention also provides a computer storage medium, wherein the computer storage medium may store a program, which, when executed, may include some or all of the steps provided in the embodiments of the present invention. The storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0098] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other medium capable of storing program code. It includes several instructions to cause a computer device (which may be a personal computer, a server, or a second device, network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0099] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0100] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or modules may be electrical, mechanical, or other forms.

[0101] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0102] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0103] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A time series prediction method based on multi-scale feature reconstruction and multi-expert fusion, characterized in that, include: Acquire multi-source time-series data containing target sequence and covariate sequence, preprocess the multi-source time-series data and construct a structured feature matrix containing multi-dimensional time features; Using three sets of one-dimensional convolutional networks with different receptive fields, multi-scale feature reconstruction is performed on the sequences in the feature matrix to obtain reconstructed feature sample sets of local, intermediate and global scales; The reconstructed feature sample set is divided into time-series slices, and the statistical features within each slice are calculated; For each scale of samples after time-series slicing, multiple prediction components are obtained through parallel branches, including: basic linear extrapolation components based on historical target sequences, linear correction components based on future covariate sequences, and nonlinear residual components based on the splicing of historical and future sequences. Based on the statistical characteristics, multiple prediction components are dynamically fused through a gating network to obtain branch prediction results at each scale; all branch results are then fused through a fully connected network to output the final prediction value.

2. The method according to claim 1, characterized in that, The target sequence is an electricity load sequence or an electricity consumption sequence; the covariate sequence includes meteorological data and time characteristic data.

3. The method according to claim 2, characterized in that, Preprocessing of multi-source time-series data and construction of a structured feature matrix containing multi-dimensional temporal features includes: The multi-source time-series data is divided into training set, validation set and test set according to time order and a predetermined ratio; Encode the timestamps of the training set to construct a multidimensional time feature containing date and period attributes; The target sequence, covariate sequence, and multidimensional time features are concatenated to generate a structured feature matrix.

4. The method according to claim 2, characterized in that, Preprocessing of multi-source time-series data and construction of a structured feature matrix containing multi-dimensional temporal features includes: Time series alignment is performed based on the sampling characteristics of each data source, and training, validation and test sets are adaptively divided according to the volatility index of time series data. Timestamp information is input into a temporal feature encoding network, which automatically extracts and outputs multidimensional temporal features that fuse periodicity and event-related characteristics. The temporal feature encoding network includes a temporal autoencoder architecture. The encoder part of the temporal autoencoder includes an attention mechanism for capturing global dependencies and a convolutional module for capturing local periodic patterns. A feature relationship graph is constructed with the target sequence and covariate sequence as nodes. The node information is aggregated through a graph neural network to generate a feature matrix that encodes the correlation between variables, which serves as the structured feature matrix. The feature relationship graph is a fully connected graph, and its edge weights are initialized based on the statistical correlation or mutual information between variables.

5. The method according to claim 1, characterized in that, Using three sets of one-dimensional convolutional networks with different receptive fields, multi-scale feature reconstruction is performed on the sequences in the feature matrix to obtain reconstructed feature sample sets at three scales: local, intermediate, and global. A three-layer, depth-increasing one-dimensional convolutional network is used, corresponding to the local, intermediate, and global scales, respectively. The first group of convolutional networks consists of single-layer convolutions and is used to extract detailed features at the local scale. The second set of convolutional networks consists of two cascaded convolutional layers and is used to extract the medium-scale periodic features. The third set of convolutional networks consists of three cascaded convolutional layers and is used to extract the global-scale trend features.

6. The method according to claim 1, characterized in that, The reconstructed feature sample set is divided into time-series slices, and statistical features within each slice are calculated, including: Using a preset fixed length as the history window, the reconstructed feature sample set is sliced ​​by sliding along the time axis with a step size of 1. For each slice, calculate statistics for the target sequence and at least one key covariate within its historical window, including at least one of the mean, standard deviation, and maximum value.

7. The method according to claim 1, characterized in that, For each scale of samples after time-series slicing, multiple prediction components are obtained through parallel branches, including: Input the historical target sequence of each scale sample into a basic linear projection layer to obtain the basic linear extrapolation component; Input the future covariate sequence of each scale sample into a covariate linear projection layer to obtain the covariate linear correction component; The historical target sequence, historical covariate sequence, and future covariate sequence of each scale sample are concatenated into a conditional sequence, which is then input into a conditional autoencoder based on a multilayer perceptron to obtain nonlinear residual components.

8. The method according to claim 1, characterized in that, Based on the aforementioned statistical characteristics, multiple prediction components are dynamically fused using a gating network to obtain branch prediction results at various scales, including: The statistical features are input into a gating network, and the gating network outputs dynamic weights corresponding to the basic linear extrapolation components, covariate linear correction components, and nonlinear residual components, respectively. After normalizing the dynamic weights, the results of the three prediction components are weighted and summed to obtain the branch prediction results at the corresponding scale.

9. The method according to claim 8, characterized in that, The gated network includes a multilayer perceptron, which maps the input statistical features to the initial weights of each prediction component through nonlinear transformation, and then normalizes them using the Softmax function to output the dynamic weights with a sum of 1.

10. A time series prediction system based on multi-scale feature reconstruction and multi-expert fusion, characterized in that, include: The data acquisition module is used to acquire multi-source time-series data containing target sequences and covariate sequences, preprocess the multi-source time-series data, and construct a structured feature matrix containing multi-dimensional time features. The feature reconstruction module is used to perform multi-scale feature reconstruction on the sequence in the feature matrix using three sets of one-dimensional convolutional networks with different receptive fields, so as to obtain three sets of reconstructed feature sample sets at local, medium and global scales. The slicing module is used to perform time-series slicing on the reconstructed feature sample set and calculate the statistical features within each slice; The branch prediction module is used to obtain multiple prediction components for samples after time-series slicing at each scale through parallel branches, including: basic linear extrapolation components based on historical target sequences, linear correction components based on future covariate sequences, and nonlinear residual components based on the splicing of historical and future sequences. The branch fusion module is used to dynamically fuse multiple prediction components through a gating network based on the statistical features to obtain branch prediction results at each scale; and to fuse all branch results through a fully connected network to output the final prediction value.

Images