A Multi-element Power Consumption Load Forecasting Method

Abstract

A multivariate power consumption load forecasting method includes: processing input multi-channel historical load data through reversible instance normalization; reconstructing the phase space of each channel sequence, mapping the one-dimensional sequence to a high-dimensional space to capture the system dynamics; separating the reconstructed features into low-frequency and high-frequency components through a frequency domain dynamic decomposition module, adaptively adjusting them using a learnable filter, and then mapping them back to the time domain; inputting the obtained low-frequency and high-frequency time-domain features into two parallel channel-aware Mamba encoders for feature extraction and fusion, with the encoders scanning along the channel dimension to capture inter-variable dependencies; performing residual connection and layer normalization with the original reconstructed features after fusion; flattening and linear mapping through a prediction head to obtain the normalized prediction result, and then performing inverse normalization to obtain the final power load forecast value at the original scale. This invention can effectively coordinate the handling of both time and channel dependencies, improving prediction accuracy and stability.

Description

Technical Field

[0001] This invention relates to the application of artificial intelligence in smart grids and power load forecasting, and in particular to a multi-dimensional power consumption load forecasting method. Background Technology

[0002] Electricity load forecasting is a key technology for smart grid operation, energy dispatch, and management. Its core objective is to accurately estimate electricity demand over a future period based on historical electricity consumption data. With the expansion of power systems and the diversification of electricity consumption patterns, multivariate time series forecasting methods have attracted widespread attention because they can simultaneously consider multiple types of electrical equipment, multiple regions, or different influencing factors. However, existing multivariate electricity load forecasting technologies still face many challenges in practical applications, mainly in the following aspects:

[0003] First, electricity consumption load data itself exhibits significant non-stationarity and scale differences between channels. Load curves for different electrical devices, regions, or time scales often display their own unique means, variances, and trends. This channel-specific scale difference can interfere with the model's extraction of common time patterns, leading to unstable model training, slow convergence, and difficulty in establishing effective dependencies between variables at different scales.

[0004] Secondly, the dynamic behavior of power systems is complex, with their time series often simultaneously containing long-term trends, periodic changes, and short-term fluctuations. Traditional time-domain analysis methods struggle to explicitly decouple and differentiate these dynamic patterns across different time scales. For example, while models based on recurrent neural networks (RNNs) or ordinary convolutional neural networks (CNNs) can capture certain time dependencies, they have limitations in simultaneously modeling long-term slow changes and short-term rapid fluctuations. They are prone to overemphasizing local noise or ignoring long-term evolution patterns, thus affecting prediction accuracy.

[0005] Furthermore, the dependencies between multivariate load sequences exhibit complex spatiotemporal coupling characteristics. Many existing models either process each channel independently before simple fusion, ignoring the potential mutual influence between channels; or they employ attention-based or ordinary convolutional mechanisms to model channel relationships, but these are computationally complex and may struggle to efficiently capture global channel dependencies when dealing with long sequences. Especially when prediction strides are long, effectively modeling and utilizing this "dual dependency" between channels (i.e., the dynamics of the temporal dimension and the correlation of the channel dimension) is a key challenge.

[0006] Furthermore, effectively extracting deep features from load data with inherent dynamic characteristics is also a challenge. Electricity load changes are driven by multiple factors such as climate, economy, and user behavior; its generation process can be viewed as the output of a dynamic system. Directly operating on the original one-dimensional time series may not fully reveal the underlying deterministic structure or system state information, limiting the model's representational capabilities.

[0007] In addition, existing end-to-end prediction models often sacrifice interpretability and computational efficiency in pursuit of high accuracy. Some complex model architectures have a large number of parameters, high training costs, and their internal decision-making processes lack intuitive explanation, making them unsuitable for deployment and trust in real power systems.

[0008] Therefore, there is an urgent need for a multivariate power consumption load forecasting method that can effectively handle multi-channel scale differences and non-stationarity, explicitly decouple and model long- and short-term dynamics, efficiently capture complex inter-channel dependencies, and at the same time take into account prediction accuracy, stability and computational efficiency.

[0009] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0010] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a multi-dimensional power consumption load prediction method.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A multi-source power consumption load forecasting method includes the following steps:

[0013] S1. Perform reversible instance normalization on the input multi-channel historical power consumption load data to eliminate the scale differences and non-stationarity effects between channels.

[0014] S2. The phase space is reconstructed for each channel's power consumption load sequence after normalization, and the one-dimensional time series is mapped to a high-dimensional phase space to obtain the reconstruction features containing time and structural dimensions.

[0015] S3. The reconstructed features are processed by the frequency domain dynamic decomposition module, wherein the reconstructed features are converted to the frequency domain and separated into low-frequency components and high-frequency components. The bias of the two types of components is dynamically adjusted by an independent learnable filter. Finally, the adjusted components are mapped back to the time domain to obtain low-frequency time domain features and high-frequency time domain features respectively.

[0016] S4. Input the low-frequency time-domain features and high-frequency time-domain features into two parallel channel-aware Mamba encoders for feature extraction and fusion. The channel-aware Mamba encoder captures the inter-channel dependencies through a state-space model that scans along the channel dimension. After fusing the outputs of the two encoders, the fused features are then subjected to residual connection and layer normalization processing to obtain the fused deep features.

[0017] S5. Flatten and linearly map the fused deep features through the prediction head to obtain the multi-channel future power load prediction results under the normalized scale. Then, perform inverse normalization on the prediction results to obtain the final power consumption load prediction value under the original scale.

[0018] A computer program product includes a computer program that, when executed by a processor, implements the multi-source power consumption load prediction method.

[0019] The present invention has the following beneficial effects:

[0020] This invention proposes a multivariate power consumption load forecasting method. By introducing a Reversible Instance Normalization (RevIN) mechanism, it effectively eliminates the scale differences and non-stationarity effects between different channels in multivariate power load data. This operation not only stabilizes the model training process but also ensures that feature extraction focuses on the inherent temporal dynamic patterns of the data rather than irrelevant numerical scales. Furthermore, its reversibility guarantees that the prediction results can be losslessly mapped back to the original physical dimensions, balancing the advantages of standardization with the interpretability of the final output.

[0021] By combining phase space reconstruction and frequency domain dynamic decomposition, this invention can deeply explore the deterministic dynamic structure inherent in load sequences and explicitly decouple long-term trends from short-term fluctuations. Phase space reconstruction maps a one-dimensional time series to a high-dimensional space, preserving the system's potential dynamic information; while frequency domain decomposition based on learnable filters can adaptively enhance key frequency components and suppress noise, enabling the model to specifically handle change patterns at different time scales. This significantly improves the modeling ability for complex power load curves, especially sequences with trend, periodic, and random fluctuation characteristics, effectively capturing and utilizing dynamic dependencies in the time dimension.

[0022] To address the complex inter-channel dependencies in multivariate load forecasting, this invention's channel-aware Mamba encoder achieves efficient capture of cross-variable dependencies through a state-space model that scans along the channel dimension. This design abandons the high computational overhead of traditional attention mechanisms, modeling long-range channel correlations with linear complexity, thereby achieving efficient modeling of both temporal and channel dependencies. By combining parallel processing of low-frequency and high-frequency branches with residual connections to the original features, the model can learn deep, discriminative corrected features while preserving the basic phase space information, enhancing the robustness and discriminative power of the feature representation.

[0023] Furthermore, the entire methodology embodies an efficient and lightweight design. The channel-shared weight prediction head significantly reduces the number of model parameters, lowers the risk of overfitting, and improves training stability and generalization ability. From reversible normalization and frequency domain dynamic adaptive decomposition to the efficient channel-aware scanning mechanism, the components work collaboratively, enabling this invention to ensure prediction accuracy while also possessing good computational efficiency and adaptability to different prediction scenarios. This provides an accurate, reliable, and practical solution for load forecasting tasks in practical power systems.

[0024] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0025] Figure 1 This is a flowchart of the overall process for predicting the multi-source power consumption load of the present invention.

[0026] Figure 2 This is an algorithm architecture diagram of the multi-element power consumption load prediction method according to an embodiment of the present invention.

[0027] Figure 3 This is a prediction effect diagram of the multi-element power consumption load prediction method according to an embodiment of the present invention. Detailed Implementation

[0028] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0029] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0030] This invention aims to address the challenges of data non-stationarity, long-term and short-term dynamic coupling, and complex dependencies between channels in multivariate power load forecasting. It proposes a multivariate power consumption load forecasting method and establishes an end-to-end forecasting framework that integrates reversible instance normalization, phase space reconstruction, frequency domain dynamic decomposition, and channel-aware state space modeling. This framework can adaptively eliminate scale differences, explicitly decouple and model long-term trends and short-term fluctuations separately, while efficiently capturing the spatiotemporal dependencies between multivariate sequences with linear computational complexity. Thus, it improves the stability, generalization ability, and computational efficiency of the model while ensuring forecast accuracy.

[0031] See Figure 1 and Figure 2 This invention provides a method for predicting multi-source power consumption load, comprising the following steps:

[0032] Step S1: Perform reversible instance normalization on the input multi-channel historical power consumption load data to eliminate the scale differences and non-stationarity effects between channels.

[0033] In some embodiments, step S1, the reversible instance normalization process specifically includes: for the power consumption load sequence of each channel, calculating its mean and standard deviation over historical time steps; and using the mean and standard deviation to standardize the time series of the channel. In the subsequent step S5, the prediction results are subjected to an inverse linear transformation using the mean and standard deviation saved for each channel during training to restore the original data scale.

[0034] Step S2: Perform phase space reconstruction on the normalized power consumption load sequence of each channel, and map the one-dimensional time series to a high-dimensional phase space to obtain the reconstruction features containing time dimension and structural dimension.

[0035] In some embodiments, step S2 specifically includes: setting a time delay parameter and an embedding dimension parameter for each channel; for each time point, selecting a set of historical observations arranged by time delay intervals from the power consumption load sequence of that channel according to the time delay parameter and the embedding dimension parameter, and constructing a high-dimensional embedding vector; the embedding vectors of all time points together constitute the reconstruction feature of that channel; the reconstruction features of all channels together form the reconstruction feature containing the time dimension and the structural dimension.

[0036] Step S3: The reconstructed features are processed by the frequency domain dynamic decomposition module, specifically including: converting the reconstructed features to the frequency domain and separating them into low-frequency components and high-frequency components, then dynamically adjusting the bias of the two types of components through independent learnable filters, and finally mapping the adjusted components back to the time domain to obtain low-frequency time domain features and high-frequency time domain features respectively.

[0037] In some embodiments, step S3 specifically includes the following processing by the frequency domain dynamic decomposition module:

[0038] S31. Perform a fast Fourier transform on the reconstructed features along their time dimension to obtain their frequency domain representation;

[0039] S32. Sort the frequency components according to their amplitude values ​​in the frequency domain representation, select the first predetermined number of frequency components with the largest amplitude as low-frequency components, and the remaining components as high-frequency components.

[0040] S33. Input the low-frequency component and the high-frequency component into two independent learnable one-dimensional convolutional filters respectively. The filters dynamically adjust the bias of the input components through their weight matrix and bias vector, and output the adjusted frequency domain components after passing through the activation function.

[0041] S34. Perform inverse fast Fourier transform on the dynamically adjusted low-frequency and high-frequency frequency domain components respectively, and convert them back to the time domain to obtain the low-frequency time domain features and high-frequency time domain features.

[0042] Step S4: Input the low-frequency time-domain features and high-frequency time-domain features into two parallel channel-aware Mamba encoders for feature extraction and fusion. The channel-aware Mamba encoders capture the inter-channel dependencies through a state-space model that scans along the channel dimension. After fusing the outputs of the two encoders, the fused features are then subjected to residual connection and layer normalization processing to obtain the fused deep features.

[0043] In some embodiments, step S4, the processing of the channel-aware Mamba encoder specifically includes:

[0044] S41. Enhance the feature dimension of low-frequency or high-frequency time-domain features of the input through a linear projection layer;

[0045] S42. The projected features are input into the state space model along the channel dimension for sequential scanning. The state space model recursively calculates the hidden state and output of the current channel based on the projected features of the current channel and the hidden state of the previous channel.

[0046] S43. The final outputs of the two encoders that process low-frequency and high-frequency features respectively are fused.

[0047] S44. Add the residuals of the fused features and the original reconstructed features obtained in step S2, and perform layer normalization on the addition result to obtain the fused deep features.

[0048] In some embodiments, in step S43, the fusion is to add the outputs of the low-frequency branch and the high-frequency branch element by element; if the size of the added feature in the structural dimension is different from the embedding dimension of the reconstructed feature in step S2, then it is mapped to the same dimension through a linear projection layer before performing the residual connection in step S44.

[0049] Step S5: Flatten and linearly map the fused deep features through the prediction head to obtain the multi-channel future power load prediction results under the normalized scale. Then, perform inverse normalization on the prediction results to obtain the final power consumption load prediction value under the original scale.

[0050] In some embodiments, step S5, the processing of the prediction head specifically includes:

[0051] S51. For the fused deep features, flatten the features of each channel in the time dimension and structural dimension into a one-dimensional feature vector.

[0052] S52. Use a linear mapping layer with shared weights between channels to map the flattened one-dimensional feature vector of each channel into a power load prediction sequence for multiple future time steps.

[0053] S53. Using the mean and standard deviation saved for each channel in step S1, perform an inverse transformation on the prediction sequence under the normalized scale obtained by the linear mapping to obtain the final power consumption load prediction value of each channel under the original scale.

[0054] In some embodiments, the input multi-channel historical time series data is multivariate historical power load data, and the final prediction result is a multivariate power load prediction value for multiple future time steps.

[0055] In some embodiments, the parameters for phase space reconstruction, frequency domain decomposition, number of channel-aware Mamba encoder layers, and model training parameters are set and optimized according to the prediction task.

[0056] The multivariate power consumption load forecasting method proposed in this invention has the following main technical advantages: it effectively eliminates the scale differences and non-stationarity of multi-channel data through reversible instance normalization; it explicitly separates and adaptively processes long- and short-term dynamic features using phase space reconstruction and frequency domain dynamic decomposition to capture time-dimensional dependencies; and it employs a channel-aware Mamba encoder to efficiently capture the dependencies between multivariate variables with linear computational complexity, thereby modeling channel-dimensional dependencies. This method systematically and collaboratively handles both time and channel dependencies, achieving high accuracy, strong stability, good interpretability, and computational efficiency in multivariate load forecasting.

[0057] The following further describes specific embodiments of the present invention, algorithm examples, and experimental verification.

[0058] A multivariate power consumption load forecasting method based on phase space reconstruction and dual dependencies is proposed. A power consumption load forecasting framework is constructed, with the core forecasting target being the power load value. This framework comprises five core components: Reversible Instance Normalization (RevIN), Phase Space Reconstruction (PSR), Frequency Domain Dynamic Decomposition Module, Channel-Aware Mamba Encoder, and a forecast head with inverse normalization. The overall architecture is as follows: Figure 2 As shown, the specific implementation methods of each component are as follows.

[0059] Problem definition:

[0060] Given a historical electricity consumption load dataset ,in This indicates the number of channels (i.e., the number of variables, the number of electrical devices). This represents the input time step, which is also the historical time step. The research objective is to learn a mapping function. To achieve all The future of each channel The prediction is performed at each time step (prediction step size) to obtain the prediction result. .

[0061] Reversible instance normalization (RevIN):

[0062] Before feature extraction, Reversible Instance Normalization (RevIN) is applied to the input historical power consumption load data to reduce the scale differences and non-stationarity effects between different channels. This is crucial for stabilizing the training process and improving prediction accuracy.

[0063] Operation process: For each channel The normalized calculation is as follows: in, and They represent the first The mean and standard deviation of each channel. It is a very small constant used to avoid the case where the denominator is zero.

[0064] The Reversible Instance Normalization (RevIN) design effectively addresses the challenges of channel-specific scale differences and non-stationarity commonly found in electricity consumption load data. This mechanism effectively eliminates scale bias between different variables by independently normalizing each channel. Its reversible nature ensures that the results can be losslessly mapped back to the original data scale through inverse transformation in the final prediction stage, maintaining information integrity. Simultaneously, through centralization and standardization operations, RevIN can partially mitigate non-stationary factors such as trends and periodicity in the data, allowing the model to focus more on learning the inherent temporal dynamics of the data, rather than being disturbed by irrelevant scale differences, thereby improving the model's training stability and prediction accuracy.

[0065] Phase Space Reconstruction (PSR):

[0066] To capture the underlying dynamic structure of the power consumption system, delayed embedding is performed on each channel based on Takens' embedding theorem, and the normalized data is processed. Phase space reconstruction is performed to uncover the deterministic structure hidden in the original time domain within multivariate time series.

[0067] For each channel time series Transformation to a high-dimensional phase space is achieved through delayed embedding. Specifically, a time delay is chosen. With Embedded Dimension For each time step Construct the embedding vector:

[0068] Among them, time delay The embedding dimension is determined using the mutual information method. Feature maps are selected using the False Nearest Neighbor (FNN) algorithm. After reconstruction, the output feature map is... ,in This indicates the length of the sequence after delayed embedding (i.e., the embedding time step).

[0069] Phase space reconstruction (PSR) is designed based on the implicit dynamic characteristics of power consumption load systems. Because power data exhibits complex interdependencies between channels, and its temporal dynamics are often insufficiently represented in the original one-dimensional sequence, PSR maps the data to a high-dimensional space through delayed embedding, thereby preserving the potential attractor structure of the system behavior. (Embedding Dimension) Used to capture the minimum variable information needed to describe the system state, while time delay This ensures that the constructed embedding vector contains non-redundant temporal evolution information. The final output feature map... Integrating the time dimension (across) ) and structural dimensions (across) The information is used to form a richer feature representation, which provides strong support for subsequent frequency domain decomposition and feature extraction.

[0070] Frequency domain dynamic decomposition module:

[0071] Features after reconstructing the phase space The input frequency domain dynamic decomposition module first separates the data into low-frequency (LF) and high-frequency (HF) components, and then dynamically adjusts the bias of the two types of components through a learnable filter to enhance the model's ability to capture long-term trends and short-term fluctuations.

[0072] Frequency separation based on FFT:

[0073] First of all, Embedded time step dimension ( Perform a Fast Fourier Transform (FFT) to transform the data from the time domain to the frequency domain. For each channel... With Embedded Dimension The FFT operation is as follows:

[0074] in, Indicates the first The first channel, the first The length is in each embedding dimension. sequence.

[0075] Then, sort by frequency component amplitude (absolute value) in descending order and select the top... Each component serves as low-frequency information (LF), corresponding to long-term trends, periodicity, or slowly changing dynamic characteristics; the remaining components... Each component is used as high-frequency (HF) information, representing short-term fluctuations, noise, or rapidly changing dynamic characteristics, and is typically taken as... or .

[0076] Dynamic bias adjustment based on learnable filters:

[0077] To adaptively enhance or suppress key frequency components, two independent learnable filters are designed, one for low-frequency components and one for high-frequency components. Each filter is a one-dimensional convolutional layer with a kernel size of 1 (to avoid changing the frequency dimension), and the activation function is LeakyReLU, defined as follows:

[0078] in, This is the weight matrix. For bias vectors, This indicates the frequency domain characteristics of low-frequency (LF) or high-frequency (HF).

[0079] The core function of this filter is to dynamically adjust the bias of frequency components based on data characteristics, specifically as follows:

[0080] For low-frequency components, the weights of frequency components related to the dominant trend are increased, while irrelevant low-frequency noise is suppressed to ensure that the model focuses on stable long-term patterns.

[0081] For high-frequency components, meaningful short-term fluctuations (such as abrupt changes in sensor data) are adaptively preserved while random noise is reduced to avoid overfitting the model to high-frequency noise.

[0082] iFFT-based temporal mapping backhaul:

[0083] After dynamic bias adjustment, the filtered low-frequency and high-frequency domain features are subjected to inverse fast Fourier transform (iFFT) to map them back to the time domain, resulting in two time-domain feature maps:

[0084] in, and These represent low-frequency and high-frequency characteristics in the time domain, respectively.

[0085] The frequency domain dynamic decomposition module effectively decouples and enhances the dynamic characteristics of power load sequences at different time scales. Through frequency domain separation, this module explicitly distinguishes between low-frequency components characterizing long-term trends and high-frequency components reflecting short-term fluctuations, allowing subsequent models to specifically address these two significantly different dynamic patterns. Compared to traditional fixed-frequency filters, the learnable filter introduced in this module possesses adaptive adjustment capabilities, dynamically enhancing key frequency components and suppressing noise based on the characteristics of the input data, thereby improving the model's generalization performance across different power consumption patterns. Simultaneously, the frequency-to-time domain conversion mechanism based on the Fast Fourier Transform (FFT) and its inverse (iFFT) fully preserves the temporal structure of the data while maintaining computational efficiency, ensuring that the decomposed features have both clear physical interpretability and remain correlated with the original sequence.

[0086] Channel-aware Mamba encoder:

[0087] The decomposed low-frequency and time-domain high-frequency features and Two parallel channel-aware Mamba encoders are input to extract deep temporal features and inter-channel correlation features. Compared to the original Mamba, the improved Mamba has two key adjustments: removing the depthwise convolutional module and setting the scan dimension to the channel dimension. ).

[0088] Improved Mamba layer design:

[0089] Each Mamba encoder is composed of It consists of stacked, improved Mamba layers. For a single layer, the input is a feature map. The main processing steps are as follows:

[0090] Linear projection: embedding dimensions through linear layers Mapping to intermediate dimensions (usually taken) The calculation is as follows: .

[0091] Channel-level scan: Unlike the original Mamba (scanning along the time dimension), this scan follows the channel dimension to capture the dependencies between variables. The scan operation is implemented using a state-space model (SSM), with the following recursive relationship: Where A, B, C, and D are learnable SSM parameters. For the first The hidden state of each channel.

[0092] Fusion of low-frequency and high-frequency features:

[0093] go through After processing by the layer-modified Mamba encoder, the low-frequency and high-frequency branches are obtained respectively: in This represents the internal dimensions of the Mamba layer. While ensuring dimension additivity, branch merging is performed: like Introducing a linear projection layer Map H back to phase space reconstruction Same embedding dimension:

[0094] like Therefore, a linear projection layer is not required: .

[0095] Residual connection, normalization, and output representation:

[0096] The fused features Original features obtained from phase space reconstruction Perform residual connections: The channel-aware Mamba encoder design utilizes parallel feature extraction paths to process low-frequency components representing long-term trends and high-frequency components reflecting short-term fluctuations, enabling the model to adapt to two types of dynamic modes using different parameterization capabilities. Feature fusion employs element-wise summation, preserving information superposition while simplifying the subsequent prediction head structure. If the feature dimensions after Mamba layer transformation are inconsistent with the original phase space features, linear projection is used to align the dimensions, ensuring the executability and numerical stability of residual connections. This residual structure retains the original features obtained from phase space reconstruction as basic information, allowing the Mamba encoder output to serve only as a learnable correction. Layer normalization stabilizes the feature distribution, jointly promoting the convergence speed and generalization performance of model training.

[0097] Prediction Head (Flatten, Linear Mapping, and Inverse Normalization):

[0098] Flattening vs. Linear Prediction:

[0099] To map 3D features to The prediction target is determined using a channel-shared linear prediction head. First, for each channel... Will Flatten: Then, the prediction is obtained through a linear mapping of channel sharing: The design of the prediction head, with its channel-sharing linear mapping, not only achieves global coupling and synergy between predictions of different channels, but also significantly reduces the number of model parameters, thereby enhancing the stability and efficiency of the training process.

[0100] Reverse RevIN:

[0101] Let the mean and standard deviation of each channel be stored during training. Prediction under normalized scale Perform inverse transform: The entire output is This serves as a forecast of future power consumption load for all channels.

[0102] The inverse normalization step utilizes the mean and standard deviation parameters saved for each channel during the training phase to perform an inverse transformation on the normalized scale prediction value output by the model, mapping it back to the original data dimensions, so that the final power load prediction result has clear physical meaning and practical interpretability.

[0103] The training steps and parameters for the power load forecasting model are as follows.

[0104] Dataset and training metrics:

[0105] The ECL dataset, short for Electricity Consuming Load, is a classic multivariate dataset commonly used in time series forecasting for electricity demand forecasting tasks. It is frequently used to validate the performance of long-term and short-term forecasting models. The dataset contains electricity consumption data from 321 customers (or terminals). The experiments involve predicting electricity load data for the next 96, 192, 336, and 720 steps with an input step size of 96. The training metrics are Mean Absolute Error (MAE) and Mean Squared Error (MSE).

[0106] Training steps:

[0107] The model's data flow can be divided into five consecutive processes: Reversible Instance Normalization (RevIN), Phase Space Reconstruction (PSR), Frequency Domain Dynamic Decomposition Module, Channel-Aware Mamba Encoder, and Prediction Head with Inverse Normalization. The delay embedding dimension in phase space reconstruction is set to 8, and the delay is set to 2; k is set to 24 in the frequency domain decomposition process; the number of layers in the channel-aware Mamba encoder is... The learning rate was set to 2; the learning rate took values ​​in the range {0.0005, 0.001, 0.002} depending on the prediction step size, and the batch size was set to 8. Training and testing were performed on a single server equipped with an NVIDIA GeForce 3090 GPU, an Intel Core i9-11900k CPU, and 32GB of DDR4 4000MHz memory. The software used was the PyTorch 1.13.0 framework and the CUDA Toolkit 11.6 environment.

[0108] Experimental prediction results are as follows Figure 3 As shown. By Figure 3 As can be seen, the overall fluctuation trend of the predicted value is highly consistent with that of the actual value. The predicted value is closely aligned with the actual value at most time points, with only slight deviations in a few local intervals where the fluctuations are more drastic. It can well reproduce the dynamic change characteristics of the actual value and demonstrates the good prediction and fitting effect of this method.

[0109] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0110] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0111] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0112] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc or CD-ROM; magnetic surface memory can be disk storage or magnetic tape storage. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0113] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0114] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0115] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0116] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this 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 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0118] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0119] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0120] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0121] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A multi-element power consumption load forecasting method, characterized in that, Includes the following steps: S1. Perform reversible instance normalization on the input multi-channel historical power consumption load data to eliminate the scale differences and non-stationarity effects between channels. S2. The phase space is reconstructed for each channel's power consumption load sequence after normalization, and the one-dimensional time series is mapped to a high-dimensional phase space to obtain the reconstruction features containing time and structural dimensions. The phase space reconstruction specifically includes: setting a time delay parameter and an embedding dimension parameter for each channel; for each time point, selecting a set of historical observations arranged by time delay intervals from the power consumption load sequence of that channel according to the time delay parameter and the embedding dimension parameter, and constructing a high-dimensional embedding vector; the embedding vectors of all time points together constitute the reconstruction feature of that channel; the reconstruction features of all channels together form the reconstruction feature containing the time dimension and the structural dimension. S3. The reconstructed features are processed by a frequency domain dynamic decomposition module, wherein the reconstructed features are converted to the frequency domain and separated into low-frequency components and high-frequency components. Then, the biases of the two types of components are dynamically adjusted by independent learnable filters. Finally, the adjusted components are mapped back to the time domain to obtain low-frequency time domain features and high-frequency time domain features, respectively. The low-frequency components and the high-frequency components are respectively input into two independent learnable one-dimensional convolutional filters. The filters dynamically adjust the biases of the input components through their weight matrices and bias vectors, and output the adjusted frequency domain components after passing through an activation function. S4. Input the low-frequency time-domain features and high-frequency time-domain features into two parallel channel-aware Mamba encoders for feature extraction and fusion. The channel-aware Mamba encoder captures the inter-channel dependencies through a state-space model that scans along the channel dimension. After fusing the outputs of the two encoders, the fused features are then subjected to residual connection and layer normalization processing to obtain the fused deep features. S5. Flatten and linearly map the fused deep features through the prediction head to obtain the multi-channel future power load prediction results under the normalized scale. Then, perform inverse normalization on the prediction results to obtain the final power consumption load prediction value under the original scale.

2. The method as described in claim 1, characterized in that, In step S1, the reversible instance normalization process specifically includes: for the power consumption load sequence of each channel, calculating its mean and standard deviation at historical time steps; using the mean and standard deviation to perform a standardization calculation on the time series of the channel; and in step S5, using the mean and standard deviation saved for each channel during the training process, performing an inverse linear transformation on the prediction results to restore them to the original data scale.

3. The method as described in claim 1, characterized in that, In step S3, the processing of the frequency domain dynamic decomposition module specifically includes: Perform a Fast Fourier Transform along the time dimension on the reconstructed features to obtain their frequency domain representation; The frequency components are sorted according to their amplitude values ​​in the frequency domain representation, and the first predetermined number of frequency components with the largest amplitudes are selected as low-frequency components, while the remaining components are selected as high-frequency components. The low-frequency and high-frequency components after dynamic adjustment are subjected to inverse fast Fourier transforms to convert them back to the time domain, thereby obtaining the low-frequency time domain features and high-frequency time domain features.

4. The method as described in claim 1, characterized in that, In step S4, the processing of the channel-aware Mamba encoder specifically includes: The feature dimension of low-frequency or high-frequency time-domain features of the input can be increased by using a linear projection layer; The projected features are input into the state space model along the channel dimension for sequential scanning. The state space model recursively calculates the hidden state and output of the current channel based on the projected features of the current channel and the hidden state of the previous channel. The final outputs of the two encoders, which process low-frequency and high-frequency features respectively, are fused together. The fused features are added to the original reconstructed features obtained in step S2 by residual addition, and the addition result is subjected to layer normalization to obtain the fused deep features.

5. The method as described in claim 4, characterized in that, The fusion involves adding the outputs of the low-frequency branch and the high-frequency branch element by element. If the size of the added feature in the structural dimension is different from the embedding dimension of the reconstructed feature in step S2, it is mapped to the same dimension through a linear projection layer before residual connection is performed.

6. The method as described in claim 1, characterized in that, In step S5, the processing of the prediction head specifically includes: For the fused deep features, the features of each channel in the time and structural dimensions are flattened into a one-dimensional feature vector; A linear mapping layer with shared weights between channels is used to map the flattened one-dimensional feature vector of each channel into a power load prediction sequence for multiple future time steps. Using the mean and standard deviation saved for each channel in step S1, the prediction sequence under the normalized scale obtained by the linear mapping is inversely transformed to obtain the final power consumption load prediction value of each channel under the original scale.

7. The method as described in claim 1, characterized in that, The input multi-channel historical time series data is multivariate historical power load data, and the final prediction result is the multivariate power load prediction value for multiple future time steps.

8. The method according to any one of claims 1 to 7, characterized in that, The parameters for phase space reconstruction, frequency domain decomposition, number of channel-aware Mamba encoder layers, and model training parameters are set and optimized according to the prediction task.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-source power consumption load prediction method as described in any one of claims 1 to 8.

Images