A power load prediction method, system and medium based on multi-component coding

Abstract

The application discloses a kind of power load prediction method, system and medium based on multi-component coding.The method extracts three components of trend, daily seasonality and weekly seasonality by multi-seasonal MSTL decomposition;Adopt the coding structure of shared backbone and component adapter, encode each STL component and 39-dimensional external features (including load history, temperature, humidity, wind speed, time cycle, calendar and business feature groups) respectively;Through the direct output of future 96 steps (24 hours) of load prediction of gated attention fusion network and hierarchical multi-scale MIMO regression head.The experimental results on NYISO public data set show that, compared with single-path LSTM model and single-path Transformer model, the method of the application improves MAE and MAPE indicators, the error accumulation of long-term multi-step prediction is inhibited, the parameter quantity is reduced by about 40-60%, and is suitable for short-term and medium-term load prediction tasks of power system.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, specifically a power load forecasting method, system, and medium based on multi-component coding. Background Technology

[0002] Electricity load forecasting is a fundamental task in power system dispatching and planning. Accurate load forecasting helps ensure the safe operation of the power system, optimize power generation plans, and reduce operating costs.

[0003] Traditional methods mainly include time series analysis (such as ARIMA) and regression analysis, but they have limitations in handling complex nonlinear relationships and multi-scale time dependencies, making it difficult to uncover deep features in load data. In recent years, neural network-based prediction methods have gradually become a research hotspot, but existing technologies still have the following problems:

[0004] 1. Lack of multi-scale feature separation mechanism

[0005] Existing methods typically use a single neural network (such as LSTM, GRU, or Transformer) to directly model the original load sequence, failing to separate features at different time scales within the sequence. Electricity load simultaneously exhibits long-term trends, daily seasonality (24-hour cycle), and weekly seasonality (7-day cycle), among other patterns, which differ significantly in statistical properties. Directly modeling the original sequence makes it difficult for the model to simultaneously capture multi-scale time dependencies, resulting in a mixture of long-term trends and short-term periodic fluctuations, thus impacting prediction accuracy.

[0006] 2. A single encoder causes information interference.

[0007] Existing methods use a single encoder to process all input features (including trend, seasonal, and external features), failing to model the characteristics of different components separately. Different components differ significantly in their statistical distribution and variation patterns—trend components change slowly, seasonal components fluctuate periodically, and external features are discretized—using the same encoder leads to information interference between different components, making it difficult for the encoder to learn targeted feature representations. Furthermore, a completely independent encoder design results in an excessive number of parameters and low training efficiency.

[0008] 3. Insufficient utilization of external features

[0009] Existing methods handle external features (such as meteorological data, temporal features, and operational rules) in a relatively simple manner, typically employing simple concatenation or linear transformations, failing to extract semantic information from these external features. There is a complex nonlinear relationship between external features and load; simple feature engineering cannot capture these influences, especially under special scenarios such as extreme weather and holidays, where prediction accuracy degrades.

[0010] 4. Accumulation of multi-step prediction errors

[0011] Existing methods often employ a recursive approach in multi-step prediction—predicting the first step, then using the prediction result as input to predict the second step, and so on. This approach leads to a gradual accumulation of errors, especially when the number of prediction steps is long (e.g., 24 hours), where the prediction accuracy decreases in later time steps, resulting in poor smoothness of the overall prediction curve.

[0012] Defect 5: Single integration mechanism

[0013] Existing methods, when fusing multiple components or features, typically employ simple concatenation or fixed-weighted averaging, lacking an adaptive fusion mechanism. The importance of different components varies at different times; for example, during periods of rapid load change, trend components and short-term lag features may be more important, while during periods of stability, long-term seasonal features may be more important. Fixed-weight fusion cannot dynamically adjust the contribution of each component based on specific time points.

[0014] To address the aforementioned issues, this invention proposes a power load forecasting method based on MSTL decomposition, shared backbone and component adapters, grouped feature encoding, gated attention fusion, and hierarchical multi-scale MIMO regression. This method solves problems in existing technologies such as insufficient feature extraction, interference from multiple component information, inadequate utilization of external features, accumulation of multi-step prediction errors, and a single fusion mechanism. Summary of the Invention

[0015] The technical problem to be solved by this invention is: how to decompose the features of different time scales in the power load sequence, improve the prediction accuracy by sharing the parameters of the backbone and component adapters, grouping external feature encoding, gated attention adaptive fusion and hierarchical multi-scale MIMO regression mechanism, and overcome the problems of multi-scale feature mixing, single encoder information interference, insufficient utilization of external features, accumulation of multi-step prediction errors and single fusion mechanism in the existing technology.

[0016] To address the above problems, this invention provides a power load forecasting method based on multi-component coding, comprising the following steps:

[0017] S1: Perform multi-seasonal time series decomposition on the original power load time series, and extract trend components, daily seasonal components and weekly seasonal components according to preset daily and weekly cycle parameters;

[0018] S2: The trend component, daily seasonal component, and weekly seasonal component extracted in step S1 are encoded using a shared coding structure. The shared coding structure includes: a shared time series modeling backbone network for extracting the time dependency features of each component; adapters set for different components for performing specific transformations on each component; a load history feature encoder for encoding the load history features constructed from the original power load time series; and a grouped external feature encoder for grouping and encoding at least one of temperature, humidity, wind speed, time cycle, calendar, and business features. The shared coding structure outputs multiple component context vectors.

[0019] S3: The multiple component context vectors output in step S2 are fused through an adaptive fusion network, which includes: a gating network for calculating the importance weights of each component; and an attention network for modeling the interaction relationships between components; and outputs a global context vector after fusion.

[0020] S4: Input the global context vector output in step S3 into the hierarchical multi-scale regression module. The regression module divides the prediction window into multiple time-scale sub-windows, outputs the prediction results of each scale step by step, and splices them together to form a complete multi-step prediction sequence.

[0021] Furthermore, in step S1, the period configuration for multi-seasonal time series decomposition is 96 points per day and 672 points per week; for large-scale historical data, each component is pre-calculated offline and cached, and the decomposition results are directly loaded from the cache during online prediction.

[0022] Furthermore, the encoding process of the shared time series modeling backbone network in step S2 includes: processing the input sequences of each component through a shared LSTM layer to output a hidden state sequence; performing self-attention calculation on the hidden state sequence through a shared multi-head attention layer; weighting and summarizing the hidden states along the time dimension through an attention pooling layer to obtain a fixed-length context vector; setting lightweight component adapters on the input and output sides of the shared backbone for different components, performing component-specific nonlinear transformations on the input sequences, performing component-specific affine transformations on the output context vectors, and using residual connections.

[0023] Furthermore, the encoding of the grouped external feature encoder in step S2 includes: dividing the external features into several semantic groups, including load history feature group, temperature feature group, humidity feature group, wind speed feature group, time cycle feature group, and calendar and business feature group; each feature group is encoded using an independent fully connected sub-network; the encoding results of each group are concatenated and fused through one or more fully connected layers to obtain the contextual representation of the external features.

[0024] Furthermore, the fusion process in step S3 includes: concatenating multiple context vectors along the component dimensions to form a tensor; using a gating network to globally model the component set, outputting a scalar weight vector for each component, and obtaining the gating weights of the relative importance of each component through Softmax normalization; weighting the context vectors of each component according to the gating weights to obtain the gating-weighted component set; using the gating-weighted component set as a query / key / value input to a multi-head self-attention layer to model the interaction relationship between components and output the interaction-represented components; and normalizing and pooling the attention output along the component dimensions to obtain the fused global context vector.

[0025] Furthermore, the implementation of hierarchical multi-scale regression in step S4 includes: the short-term prediction head outputs the prediction results for the next 0-6 hours; the short-term prediction results are concatenated with the global context vector and input into the medium-term prediction head to output the prediction results for the next 6-18 hours; the medium-term prediction results are concatenated with the global context vector and input into the long-term prediction head to output the prediction results for the next 18-24 hours; the prediction results of the short-term, medium-term and long-term scales are concatenated in the time dimension to form a complete multi-step prediction vector.

[0026] Furthermore, when training the power load prediction model composed of the multi-component encoding structure, the adaptive fusion network, and the hierarchical multi-scale regression module, the loss function used includes at least a mean squared error term, or also a loss term robust to outliers, a multi-scale consistency loss term, and an attention weight regularization term, and the following optimization strategies are adopted: The AdamW adaptive optimization method is used, coupled with a learning rate scheduler with a warm-up phase; gradient pruning is performed, with a maximum gradient norm of 1.0; Dropout is applied in the encoder, fusion network, and regression head, with a Dropout rate of 0.12-0.18; weight decay is used, with a weight decay coefficient of 0.01-0.02; an early stopping mechanism is adopted, monitoring the validation set loss, and stopping training early when the validation loss no longer decreases within a preset number of rounds; the batch size is 32-64, and the maximum number of training rounds is 100.

[0027] A power load forecasting system based on multi-component coding includes:

[0028] The data preprocessing module is used to decompose the original power load time series into multi-seasonal time series and extract trend components, daily seasonal components and weekly seasonal components according to preset daily and weekly period parameters.

[0029] An encoding module is used to receive the trend component, daily seasonal component, and weekly seasonal component, and to encode the trend component, daily seasonal component, and weekly seasonal component using a shared encoding structure. The shared encoding structure includes: a shared time series modeling backbone network for extracting the time dependency features of each component; adapters set for different components for performing specific transformations on each component; and an external feature encoder for grouping and encoding external features. The shared encoding structure outputs multiple component context vectors.

[0030] The fusion module is used to receive the multiple component context vectors and fuse them through an adaptive fusion network. The adaptive fusion network includes: a gating network for calculating the importance weights of each component; and an attention network for modeling the interaction relationships between the components; and outputs a global context vector after fusion.

[0031] The prediction module receives the global context vector, adopts a hierarchical multi-scale MIMO architecture, divides the prediction window into multiple time-scale sub-windows, outputs the prediction results of each scale step by step, and splices them together to form a complete multi-step prediction sequence.

[0032] Furthermore, the prediction module includes a short-term prediction head, a medium-term prediction head, and a long-term prediction head; the short-term prediction head outputs the prediction results for the next 0-6 hours; the short-term prediction results are concatenated with the global context vector and input into the medium-term prediction head to output the prediction results for the next 6-18 hours; the medium-term prediction results are concatenated with the global context vector and input into the long-term prediction head to output the prediction results for the next 18-24 hours; the prediction results of the short-term, medium-term, and long-term scales are concatenated in the time dimension to form a complete multi-step prediction vector.

[0033] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the power load forecasting method based on multi-component coding as described above.

[0034] Compared with the prior art, the present invention has the following beneficial effects.

[0035] By using MSTL decomposition, multi-scale features such as trends, daily seasonality, and weekly seasonality in the power load sequence are separated, enabling the model to model features at different time scales separately and improving the targeting of feature extraction.

[0036] By employing a shared backbone and component adapter coding structure, the number of model parameters is reduced while allowing for customized modeling of different components. Compared to a completely independent encoder design, the number of parameters is reduced by approximately 40–60%, resulting in improved training efficiency.

[0037] By combining grouped external feature encoding with gated attention fusion mechanism, the model's predictive ability in extreme weather, holiday and other scenarios is improved by utilizing information from multiple external features such as meteorology, time, calendar and business.

[0038] A hierarchical multi-scale MIMO regression architecture is adopted to directly output the prediction results for all future time steps, avoiding the error accumulation problem in recursive prediction, and improving the smoothness and consistency of predictions for different time periods through multi-scale design.

[0039] Experiments on publicly available power load datasets show that the validation set MAE of the method of this invention is about 0.13 in a scenario with 15-minute granularity, 7-day historical input, and 24-hour multi-step prediction. This has an accuracy advantage compared to single-path LSTM models and single-path Transformer models, while reducing the number of parameters by about 40–60%. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of a power load forecasting method based on multi-component coding according to the present invention.

[0041] Figure 2 This is a schematic diagram of the decomposition of the multi-seasonal STL (MSTL) of the present invention;

[0042] Figure 3 This is a schematic diagram of the encoder structure of the present invention;

[0043] Figure 4 This is a complete architecture diagram of the prediction model of this invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1: A Power Load Forecasting Method Based on Multi-Component Coding

[0046] like Figure 1 As shown, this embodiment provides a power load forecasting method based on multi-component coding. The data comes from publicly available historical power load data from NYISO, an independent operator in New York, USA, and is integrated with meteorological data and calendar information. The time granularity is 15 minutes, that is, one sampling point every 15 minutes.

[0047] 1. Data Acquisition and Preprocessing

[0048] In this embodiment, the original power load time series is derived from historical power load data publicly released by NYISO, an independent systems operator in New York, USA. The original power load time series refers to a sequence of power load data collected at fixed time intervals and arranged chronologically, denoted as Y={y1,y2,...,yN}. ,in Indicates the first The power load value at each sampling time. This embodiment uses load data with a 15-minute time granularity, that is, a sampling point is formed every 15 minutes.

[0049] After obtaining publicly available historical power load data from NYISO, the load data is sorted, deduplicated, and aligned at equal intervals based on timestamps to obtain the original power load time series with a 15-minute granularity. For missing sampling points, linear interpolation is used to fill in the missing points. For outlier sampling points, the 3σ principle is used for detection, and outliers are replaced with the average of the adjacent normal sampling points. Subsequently, the processed original power load time series is normalized to obtain the load time series used for subsequent multi-seasonal time series decomposition.

[0050]

[0051] Corresponding to the original power load time series, meteorological data and calendar information at the same timestamp are obtained. The meteorological data includes at least one of temperature, humidity and wind speed, and the calendar information includes at least one of holiday identifiers, weekend identifiers, weekday identifiers and peak business period identifiers. The meteorological data and calendar information are used to construct external features and are not the processing objects of multi-season time series decomposition in step S1.

[0052] 2. Multi-Seasonal STL (MSTL) Decomposition

[0053] like Figure 2 As shown, the preprocessed raw power load time series is decomposed into multi-seasonal STL (MSTL) components, and multiple periodic components are extracted at once:

[0054] Daily cycle: period day =96 points (24 hours × 4 points / hour);

[0055] Weekly cycle: period week =672 points (7 days × 96 points / day).

[0056] The original load sequence was decomposed using the MSTL decomposition method, yielding:

[0057] Trend Components The dimensions are the same as the original sequence;

[0058] Daily seasonal components The dimensions are the same as the original sequence;

[0059] Weekly seasonal components The dimensions are the same as the original sequence.

[0060] The quality of the decomposition can be verified through the reconstruction error:

[0061] Reconstruction error =

[0062] in, This represents the preprocessed original power load time series, which is the MSTL decomposed input sequence obtained after timestamp alignment, missing value imputation, outlier handling, and normalization of NYISO publicly available historical power load data; These represent the trend component, daily seasonal component, and weekly seasonal component obtained from MSTL decomposition, respectively. The reconstruction error is used to measure the degree to which the sum of the trend component, daily seasonal component, and weekly seasonal component restores the original electricity load time series.

[0063] The average reconstruction error in this embodiment is approximately 1.2 × 10⁻⁶. -15 The accuracy is close to that of a machine, indicating that the decomposition was correct.

[0064] 3. External Feature Extraction

[0065] While performing multi-seasonal decomposition on the load time series, this invention constructs an external feature vector f to describe meteorological conditions, temporal periodicity characteristics, and operational rules. In a preferred embodiment, the total dimension of the external feature is 39, divided into the following groups:

[0066] 1. Load history characteristics group (11 dimensions): including multi-scale lag characteristics (1 step, i.e., 15 minutes ago; 4 steps, i.e., 1 hour ago; 96 steps, i.e., the same time of the previous day; 672 steps, i.e., the same time of the previous week), multi-time window rolling mean and variance (6-hour and 24-hour rolling mean, 6-hour and 24-hour rolling variance, 7-day rolling mean), 15-minute difference and daily difference;

[0067] 2. Temperature Feature Group (10 dimensions): Includes current air temperature and multi-scale lag (1 step, 2 steps, 4 steps i.e. 1 hour, 24 steps i.e. 6 hours), multi-time window rolling mean (1 hour, 6 hours, 24 hours) and rolling variance (6 hours, 24 hours), used to characterize the impact of temperature fluctuations on air conditioning load;

[0068] 3. Humidity feature group (7 dimensions): including current humidity and 6-hour lag (24 steps), multi-time window rolling mean (6 hours, 24 hours, 7 days) and rolling variance (6 hours, 24 hours), used to characterize the effects of perceived temperature;

[0069] 4. Wind speed characteristic group (2D): includes current wind speed and 6-hour lag (24 steps), used to capture the indirect impact of weather on load;

[0070] 5. Time Cyclic Feature Group (4-dimensional): Includes hourly sine / cosine encoding (hoursin, hourcos) based on a 96-point cycle (number of sampling points per day at a 15-minute granularity), and weekday cycle encoding.

[0071]

[0072] 6. Calendar and Business Characteristics Group (5 dimensions): Includes holiday identifier (is_holiday), weekend identifier (is_weekend), weekday identifier (is_business_day), business peak identifier (is_peak_7_22, 7:00-22:00 peak), and morning peak identifier (rush_morning, 7-9:00), used to highlight the morning load peak.

[0073] The aforementioned 39 external features are uniformly constructed in the preprocessing stage, and after standardization, they are input into the group feature encoder for group encoding and fusion.

[0074] 4. Encoder Implementation

[0075] like Figure 3 As shown, this embodiment adopts a structure of "shared backbone + component adapter + group feature coding", as detailed below:

[0076] Shared STL encoder (with component adapter): The trend component t, daily seasonal component sd, and weekly seasonal component sw are modeled through a shared time series modeling backbone network (referred to as the shared backbone). The shared backbone consists of a 2-layer LSTM network (128 hidden dimensions), an 8-head Multi-head Attention layer, and an attention pooling layer to obtain a fixed-length context vector. The input and output sides of the shared backbone are respectively... / / Set up a lightweight component adapter to achieve component-specific transformation through residual connections.

[0077] Load history feature encoder: The 11-dimensional historical features generated by the 15-minute granular load are encoded separately. The daily and weekly lag features are encoded using independent fully connected sub-networks. The remaining features are merged and encoded and then fused with the output of the above sub-networks.

[0078] Grouped external feature encoder: The 39-dimensional external features are divided into 6 semantic groups (load history 11-dimensional, temperature 10-dimensional, humidity 7-dimensional, wind speed 2-dimensional, time cycle 4-dimensional, calendar and business 5-dimensional). Each feature group is encoded using an independent fully connected sub-network. Then, the encoding results of each group are concatenated and fused through a fully connected layer to obtain the external feature context vector.

[0079] In a preferred embodiment, the encoder parameter statistics are approximately as follows: approximately 300,000 parameters for the shared STL encoder backbone, approximately 30,000 for the component adapter, approximately 80,000 for the load history feature encoder, approximately 120,000 for the grouped external feature encoder, and approximately 530,000 for the total encoder parameters.

[0080] 5. Fusion Network Implementation (Based on Gating and Multi-Head Attention)

[0081] In obtaining After obtaining the context vectors of the three STL components, the context vector of the load history features, and the context vector of the external features, this invention employs a fusion network containing gate networks and multi-head attention to adaptively fuse multi-component information.

[0082] Specifically, it includes:

[0083] (1) Concatenate the above five context vectors along the component dimensions to form a tensor C=[ ], its dimension is [ ];

[0084] (2) Use a gated network to model C globally, output the weight vector g∈Rncomponents for each component, and obtain the gated weights of the relative importance of each component by Softmax normalization;

[0085] (2) Weight each component context vector according to the gating weight to obtain the gating weighted component set;

[0086] (4) Use the gated weighted component set as the input of the Query / Key / Value multi-head self-attention layer, model the interaction relationship between the components, and output the interaction-represented component;

[0087] (5) Normalize and pool the attention output in the component dimension (e.g., take the average or weighted sum) to obtain the fused global context vector, which is used as the input of the subsequent multi-step prediction head.

[0088] Through the above-mentioned gating + attention fusion mechanism, this invention can adaptively adjust the contribution of trends, daily seasonality, weekly seasonality, load history and external characteristics in the final prediction according to the load and external conditions at specific times.

[0089] 6. Implementation of Multi-Scale MIMO Regression Head

[0090] To improve the consistency and smoothness of multi-step predictions, this invention employs a hierarchical multi-scale MIMO regression head, dividing multiple time steps within a future prediction window into short-term, medium-term, and long-term scales according to different time periods, and introducing residual relationships between the scales. In a preferred embodiment:

[0091] The fused global context vector is input into the short-term prediction head, which outputs the prediction results for the next 0-6 hours (24 time steps).

[0092] By concatenating the short-term forecast results with the global context vector and inputting them into the medium-term forecast head, the forecast results for the next 6-18 hours (48 time steps) are output, which can be regarded as residual correction based on the short-term forecast.

[0093] By concatenating the intermediate-term forecast results with the global context vector and inputting them into the long-term forecast head, the forecast results for the next 18-24 hours (24 time steps) are output, which can be regarded as residual correction based on the intermediate-term forecast.

[0094] Finally, the forecast results of the short-term, medium-term and long-term scales are spliced ​​together in the time dimension to form a multi-step forecast vector of length 96, corresponding to the load forecast for the next 24 hours.

[0095] This hierarchical multi-scale design can improve the accuracy and consistency of predictions across different time periods while maintaining the smoothness of the overall prediction curve.

[0096] 7. Complete model parameter statistics

[0097] like Figure 4 As shown, in a preferred embodiment, the total number of parameters of the complete model using a shared STL encoder + component adapter + grouped feature encoding + gated attention fusion + multi-scale regression head is approximately 530,000 (of which the shared STL encoder backbone has approximately 310,000, and the component adapter and load / external feature encoding and fusion and multi-scale regression head together have approximately 220,000). Compared with the comparative model (approximately 1,360,000 parameters) that does not share a backbone and encodes each component completely independently, the number of parameters is reduced by approximately 60%, thus reducing storage and computational overhead.

[0098] 8. Model Training

[0099] In a preferred embodiment, the following training configuration is used:

[0100] Loss function: a weighted combination of mean squared error, Huber loss and multi-scale consistency loss, with mean absolute error (MAE) as the main monitoring indicator;

[0101] Optimizer: AdamW or an equivalent adaptive optimization method is used, along with a OneCycle learning rate scheduler with a warm-up phase, which adjusts the maximum learning rate to about 0.0008 after 5 epochs of warm-up;

[0102] Batch size: 32 (expandable to 64 if GPU memory allows), maximum number of training epochs: 100, early stopping patience: 15-20 epochs, minimum improvement threshold: approximately 1×10⁻⁶. -4 ;

[0103] Regularization: Dropout (0.12-0.18) and weight decay (0.01-0.02) are used in the encoder, fusion network, and multi-scale regression head to suppress overfitting;

[0104] Gradient clipping: The maximum gradient norm is set to 1.0.

[0105] Experimental results on publicly available power load datasets show that, taking the current best model configuration as an example, the method of this invention can achieve a validation set MAE of approximately 0.13 (e.g., 0.1317) in scenarios with 15-minute granularity, 7-day historical input, and 24-hour multi-step prediction. This has an accuracy advantage compared to single-path LSTM models and single-path Transformer models without multi-component decomposition. After restoring to the actual power scale, the overall mean absolute error is approximately 50-60MW, the root mean square error is approximately 70-80MW, the relative error MAPE is approximately 3-4%, and the number of parameters is reduced by approximately 40-60%.

[0106] 9. Model Evaluation

[0107] The model was evaluated on the test set, and MAE, RMSE, MAPE, and R were calculated. 2 These are some of the indicators. Example results are as follows: On a normalized scale, MAE is approximately 0.13 and RMSE is approximately 0.18; when converted to a real power scale, the corresponding overall MAE is approximately 54MW, RMSE is approximately 74MW, MAPE is approximately 3.5%, and R... 2 >0.95.

[0108] Overall performance: On a normalized scale, MAE ≈ 0.13, RMSE ≈ 0.18; on a reduced scale, overall MAE is approximately 54MW, RMSE is approximately 74MW, MAPE is approximately 3.5%, and R... 2 >0.95;

[0109] Time-based performance (in 6-hour units):

[0110]

[0111] Compared to single-path LSTM and single-path Transformer models without multi-component decomposition, the method of this invention improves both MAE and MAPE metrics, especially suppressing error accumulation in long-term multi-step prediction scenarios. Furthermore, due to the shared backbone and component adapter structure, the number of parameters is reduced by approximately 40-60% compared to a completely independent encoder design, thus improving training efficiency.

[0112] Example 2: Implementation methods with different parameter configurations

[0113] This embodiment explores the impact of different parameter scales on prediction performance and resource overhead based on the current best model architecture (shared STL encoder + component adapter + grouped feature encoding + gated attention fusion + multi-scale MIMO regression). Example configurations include:

[0114] Configuration 1 (Small Model): =96, =2, =4, =0.12, while maintaining the same input and output settings, the number of parameters is reduced by about 30% compared to the standard configuration, and the validation set MAE remains in the range of 0.13-0.14, which is suitable for scenarios that are sensitive to computing resources but do not have high accuracy requirements;

[0115] Configuration 2 (Standard Model, Recommended): =128, =2, =8, =0.15, with a total of approximately 530,000 parameters. Based on the NYISO public dataset, in scenarios with 15-minute granularity, 7-day historical input, and 24-hour multi-step prediction, the best MAE on the validation set can reach approximately 0.13 (e.g., 0.1317). This corresponds to an overall MAE of approximately 50-60MW, an RMSE of approximately 70-80MW, and a MAPE of approximately 3-4% at the actual power scale. The training time and GPU memory usage are both within an acceptable range, making it a suitable recommended configuration for production deployment.

[0116] Configuration 3 (Enhanced Model): =192, =3, =8 or 12, =0.15-0.18,

[0117] Compared to the standard configuration, the number of parameters increases by about 60-80%, training time and GPU memory usage increase, but the validation set MAE on the same dataset only improves slightly or is comparable to the standard configuration, and the overall performance remains at around 0.13, resulting in relatively low cost-effectiveness.

[0118] Considering model accuracy, training cost, and deployment resource consumption, Configuration 2 achieves a good balance between performance and efficiency, making it a recommended configuration for production environments; Configuration 1 can be used in resource-constrained or edge-side scenarios; Configuration 3 is more suitable for scenarios with additional requirements for robustness to extreme scenarios and sufficient computing resources. Example 3: Power Load Forecasting System

[0119] This invention also provides a power load forecasting system based on multi-component independent coding, comprising:

[0120] The data preprocessing module is used to decompose the original power load time series into multi-seasonal time series and extract trend components, daily seasonal components and weekly seasonal components according to preset daily and weekly period parameters.

[0121] An encoding module is used to receive the trend component, daily seasonal component, and weekly seasonal component, and to encode the trend component, daily seasonal component, and weekly seasonal component using a shared encoding structure. The shared encoding structure includes: a shared time series modeling backbone network for extracting the time dependency features of each component; adapters set for different components for performing specific transformations on each component; and an external feature encoder for grouping and encoding external features. The shared encoding structure outputs multiple component context vectors.

[0122] The fusion module is used to receive the multiple component context vectors and fuse them through an adaptive fusion network. The adaptive fusion network includes: a gating network for calculating the importance weights of each component; and an attention network for modeling the interaction relationships between the components; and outputs a global context vector after fusion.

[0123] The prediction module receives the global context vector, adopts a hierarchical multi-scale MIMO architecture, divides the prediction window into multiple time-scale sub-windows, outputs the prediction results of each scale step by step, and splices them together to form a complete multi-step prediction sequence.

[0124] This invention separates trend, daily seasonality, and weekly seasonality components using MSTL decomposition, and combines a shared backbone and component adapter coding structure to reduce the number of parameters by 40%–60% while avoiding interference from multi-component information. A grouped external feature encoding and gated attention fusion mechanism fully utilize multi-source information such as meteorological and calendar data to improve adaptability to extreme scenarios. Hierarchical multi-scale MIMO regression directly outputs multi-step predictions, effectively suppressing error accumulation. Experiments show that compared to single-path LSTM / Transformer, MAE and MAPE are significantly improved, balancing accuracy and efficiency.

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. 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 be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A power load forecasting method based on multi-component coding, characterized in that, Includes the following steps: S1: Perform multi-seasonal time series decomposition on the original power load time series, and extract trend components, daily seasonal components and weekly seasonal components according to preset daily and weekly cycle parameters; S2: The trend component, daily seasonal component, and weekly seasonal component extracted in step S1 are encoded using a shared coding structure. The shared coding structure includes: a shared time series modeling backbone network for extracting the time dependency features of each component; component adapters set for different components for performing specific transformations on each component; a load history feature encoder for encoding the load history features constructed from the original power load time series; and a grouped external feature encoder for grouping and encoding at least one of temperature, humidity, wind speed, time cycle, calendar, and business features. The shared coding structure outputs multiple component context vectors. S3: The multiple component context vectors output in step S2 are fused through an adaptive fusion network, which includes: a gating network for calculating the importance weights of each component; and an attention network for modeling the interaction relationships between components; and outputs a global context vector after fusion. S4: Input the global context vector output in step S3 into the hierarchical multi-scale regression module. The regression module divides the prediction window into multiple time-scale sub-windows, outputs the prediction results of each scale step by step, and splices them together to form a complete multi-step prediction sequence.

2. The method according to claim 1, characterized in that, In step S1, the period configuration for multi-seasonal time series decomposition is 96 points per day and 672 points per week. For large-scale historical data, each component is pre-calculated offline and cached, and the decomposition results are directly loaded from the cache during online prediction.

3. The method according to claim 1, characterized in that, The encoding process of the shared time series modeling backbone network in step S2 includes: processing the input sequences of each component through a shared LSTM layer to output a hidden state sequence; performing self-attention calculation on the hidden state sequence through a shared multi-head attention layer; weighting and summarizing the hidden states along the time dimension through an attention pooling layer to obtain a fixed-length context vector; setting lightweight component adapters on the input and output sides of the shared backbone for different components, performing component-specific nonlinear transformations on the input sequences, performing component-specific affine transformations on the output context vectors, and using residual connections.

4. The method according to claim 1, characterized in that, The encoding of the grouped external feature encoder in step S2 includes: dividing the external features into several semantic groups, including load history feature group, temperature feature group, humidity feature group, wind speed feature group, time cycle feature group, and calendar and business feature group; each feature group is encoded using an independent fully connected sub-network; the encoding results of each group are concatenated and fused through one or more fully connected layers to obtain the contextual representation of the external features.

5. The method according to claim 1, characterized in that, The fusion process in step S3 includes: concatenating multiple context vectors along the component dimensions to form a tensor; using a gating network to globally model the component set, outputting a scalar weight vector for each component, and obtaining the gating weights of the relative importance of each component through Softmax normalization; weighting the context vectors of each component according to the gating weights to obtain the gating-weighted component set; using the gating-weighted component set as a query / key / value input to a multi-head self-attention layer to model the interaction relationship between components and output the interaction-based component representation; and normalizing and pooling the attention output along the component dimensions to obtain the fused global context vector.

6. The method according to claim 1, characterized in that, The implementation of hierarchical multi-scale regression in step S4 includes: the short-term prediction head outputs the prediction results for the next 0-6 hours; the short-term prediction results are concatenated with the global context vector and input into the medium-term prediction head to output the prediction results for the next 6-18 hours; the medium-term prediction results are concatenated with the global context vector and input into the long-term prediction head to output the prediction results for the next 18-24 hours; the prediction results of the short-term, medium-term and long-term scales are concatenated in the time dimension to form a complete multi-step prediction vector.

7. The method according to claim 1, characterized in that, When training the power load prediction model composed of the multi-component encoding structure, the adaptive fusion network, and the hierarchical multi-scale regression module, the loss function used includes at least a mean squared error term, or also a loss term robust to outliers, a multi-scale consistency loss term, and an attention weight regularization term. The following optimization strategies are employed: using the AdamW adaptive optimization method in conjunction with a learning rate scheduler with a warm-up phase; performing gradient pruning with a maximum gradient norm of 1.0; applying Dropout in the encoder, fusion network, and regression head with a Dropout rate of 0.12-0.18; employing weight decay with a weight decay coefficient of 0.01-0.02; using an early stopping mechanism to monitor the validation set loss and stop training early when the validation loss no longer decreases within a preset number of rounds; using a batch size of 32-64 and an upper limit of 100 training rounds.

8. A power load forecasting system based on multi-component coding, characterized in that, include: The data preprocessing module is used to decompose the original power load time series into multi-seasonal time series and extract trend components, daily seasonal components and weekly seasonal components according to preset daily and weekly period parameters. An encoding module is used to receive the trend component, daily seasonal component, and weekly seasonal component, and to encode the trend component, daily seasonal component, and weekly seasonal component using a shared encoding structure. The shared encoding structure includes: a shared time series modeling backbone network for extracting the time dependency features of each component; adapters set for different components for performing specific transformations on each component; and an external feature encoder for grouping and encoding external features. The shared encoding structure outputs multiple component context vectors. The fusion module is used to receive the multiple component context vectors and fuse them through an adaptive fusion network. The adaptive fusion network includes: a gating network for calculating the importance weights of each component; and an attention network for modeling the interaction relationships between the components; and outputs a global context vector after fusion. The prediction module receives the global context vector, adopts a hierarchical multi-scale MIMO architecture, divides the prediction window into multiple time-scale sub-windows, outputs the prediction results of each scale step by step, and splices them together to form a complete multi-step prediction sequence.

9. The system according to claim 8, characterized in that, The prediction module includes a short-term prediction head, a medium-term prediction head, and a long-term prediction head. The short-term prediction head outputs the prediction results for the next 0-6 hours. The short-term prediction results are concatenated with the global context vector and input into the medium-term prediction head to output the prediction results for the next 6-18 hours. The medium-term prediction results are concatenated with the global context vector and input into the long-term prediction head to output the prediction results for the next 18-24 hours. The prediction results of the short-term, medium-term, and long-term scales are concatenated in the time dimension to form a complete multi-step prediction vector.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the power load forecasting method based on multi-component coding as described in any one of claims 1-7.

Images