Device-level load prediction method and system based on load component dynamic decoupling

Abstract

The invention relates to the technical field of power system load prediction, in particular to an equipment-level load prediction method and system based on load component dynamic decoupling. The method is used for solving the technical problem that a traditional single prediction model cannot adapt to different load component characteristic differences. The method comprises the following steps: firstly, dynamically decomposing an equipment load into resident, commercial, industrial and other four types of components through a time-frequency domain dual-path load composition decoupling network; then, each load component is equipped with a specially optimized expert model for targeted modeling; then, the weight contribution of each expert model is adjusted in real time based on environmental factors through a dynamic gating fusion network; meanwhile, a load component memory bank is established to solve the cold start problem of new equipment; and an expert knowledge distillation mechanism is implemented, so that the expert models mutually refer to experience. Experimental results show that the method is remarkably superior to the prior art in the aspects of comprehensive prediction precision, load sudden change adaptability and cold start capability, and is particularly suitable for equipment prediction scenes with complex and changeable load components.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, specifically to an equipment-level load forecasting method and system based on dynamic decoupling of load components. Background Technology

[0002] As power systems develop towards intelligence and precision, equipment-level load forecasting has become a fundamental supporting technology for key aspects such as power grid planning, dispatching, and demand-side management. Equipment-level load forecasting primarily aims to accurately estimate the future electricity load of power equipment such as substations and distribution lines, providing crucial information for the safe and economical operation of the power grid. Compared to system-level load forecasting, equipment-level loads exhibit greater volatility and uncertainty, and typically include a mixture of different types of load components, such as residential, commercial, and industrial loads. These load components possess significantly different electricity consumption characteristics and influencing factors, posing substantial challenges to forecasting efforts.

[0003] Looking at the development history of power load forecasting technology, early load forecasting methods mainly relied on statistical principles. In the 1970s and 80s, regression analysis methods were widely used in load forecasting, such as multiple linear regression and generalized autoregressive conditional heteroscedasticity models. These methods establish forecasting models based on the statistical relationships between historical load data and relevant factors (such as temperature, humidity, and date type). Although these methods are computationally simple and easy to implement, they struggle to effectively capture the nonlinear relationships and complex patterns in load data, resulting in limited forecasting accuracy.

[0004] Subsequently, time series analysis techniques began to emerge in the field of load forecasting. Time series methods such as ARIMA (Autoregressive Integrated Moving Average) models and exponential smoothing can better characterize the temporal correlation and periodic changes in load data. These methods are particularly suitable for short-term load forecasting scenarios, but they are still insufficient when dealing with long-term trends and the influence of various external factors. Especially for load data with obvious seasonality and trends, traditional time series methods often require seasonal-trend decomposition preprocessing, increasing the complexity of modeling.

[0005] Beginning in the 1990s, artificial intelligence and machine learning technologies were gradually applied to the field of load forecasting, bringing about significant methodological breakthroughs. Artificial neural networks, due to their powerful nonlinear fitting and adaptive learning capabilities, became one of the mainstream methods for load forecasting. Subsequently, various machine learning algorithms, such as support vector machines, random forests, and gradient boosting trees, were also introduced into the field of load forecasting, further improving forecasting performance. In recent years, the rapid development of deep learning technology has brought new opportunities to load forecasting. Deep learning models such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and convolutional neural networks (CNNs) can automatically extract multi-level features from load data, significantly improving forecasting accuracy.

[0006] Despite the continuous development of load forecasting technology, current methods still face several key issues and challenges when applied to equipment-level load forecasting: First, equipment-level loads are typically a mixture of various load components, each with significantly different electricity consumption characteristics and influencing factors. For example, residential loads are mainly affected by living habits and temperature, exhibiting obvious intraday fluctuations and seasonal variations; commercial loads are highly correlated with business hours and weekday types; and industrial loads are more influenced by production plans and economic factors, exhibiting both strong regularity and suddenness. Traditional "one-size-fits-all" forecasting methods use a single model to handle mixed loads, which cannot simultaneously adapt to the differences in characteristics among different load components, resulting in limited forecast accuracy, especially when the proportion of load components changes dynamically.

[0007] Secondly, existing load forecasting methods generally lack the ability to analyze the internal structure of mixed loads. Most methods model mixed loads of different components as a whole, ignoring the composition and dynamic characteristics of the internal load components. This "black box" approach cannot identify and utilize the characteristic patterns of different load components, nor can it provide in-depth insights into load composition, thus limiting the performance ceiling and application value of forecasting models.

[0008] Third, equipment-level load forecasting faces a significant sample imbalance problem. Power systems contain a large number of newly commissioned devices or devices with significantly changed load characteristics, which often lack sufficient historical data to support modeling. Existing methods struggle to effectively handle forecasting for data-sparse devices and lack a knowledge transfer mechanism from data-rich devices to data-sparse devices, leading to a prominent cold-start problem.

[0009] Fourth, the interpretability of load forecasting models is insufficient. While black-box models such as deep learning have strong fitting capabilities, their internal decision-making mechanisms are difficult to understand, and they lack reasonable explanations for the forecast results, limiting the credibility and acceptance of these models in actual business operations. As a critical infrastructure, the power system's decision support models need to possess a certain degree of interpretability and understandability.

[0010] Fifth, the dynamic evolution of load characteristics poses challenges to forecasting. With the rapid development of new load types such as distributed energy and electric vehicles, and the continuous changes in electricity consumption patterns, load characteristics exhibit a dynamic evolution trend. Most existing forecasting methods are based on static load characteristic assumptions, lacking the ability to adapt to dynamic changes in load characteristics and making it difficult to accurately capture emerging load patterns.

[0011] To address the multi-component mixing problem in equipment-level load forecasting, the industry has proposed some preliminary solutions. A common approach is load decomposition forecasting, which involves first decomposing the mixed load into different types of sub-loads, modeling and forecasting each separately, and then aggregating the results. However, existing load decomposition methods are mostly based on simple rules or statistical features (such as weekday type, time period division, etc.), lacking in-depth analysis of the inherent patterns of the load, resulting in limited decomposition accuracy. Another approach is ensemble learning methods, which improve overall performance by integrating the prediction results of multiple base models. However, traditional ensemble methods typically employ static weight allocation strategies, failing to dynamically adjust the ensemble weights based on the characteristics of the input samples, making it difficult to adapt to scenarios where the proportions of load components change dynamically.

[0012] Furthermore, for the prediction problem of data-sparse devices, some studies have attempted to use transfer learning and few-shot learning techniques to transfer knowledge from data-rich devices to data-sparse devices. However, these methods often perform the transfer directly at the overall load level, ignoring the differences between different load components, resulting in poor transfer performance. At the same time, existing methods lack effective mechanisms for organizing and managing load knowledge, failing to achieve the accumulation and continuous optimization of load knowledge.

[0013] In practical applications, a single power equipment (such as a substation or distribution line) often carries multiple types of loads simultaneously, such as the common situation of mixed power supply to residential areas, commercial areas, and industrial parks. The electricity consumption patterns and influencing factors of these different types of loads vary significantly, and simply modeling them as a whole will result in models that cannot accurately capture the characteristics of each type of load. Furthermore, with urban development and industrial restructuring, the proportion of load components carried by equipment is constantly changing, further exacerbating the difficulty of prediction.

[0014] In summary, existing load forecasting technologies have significant limitations when dealing with mixed loads at the equipment level, making it difficult to simultaneously meet requirements such as high accuracy, interpretability, and adaptability. There is an urgent need for a novel forecasting method that can automatically identify and decompose load components, perform personalized modeling for the characteristics of different load components, and dynamically fuse the forecasting results of each component. This would overcome the limitations of traditional "one-size-fits-all" forecasting methods and improve the accuracy and practicality of equipment-level load forecasting. Based on an in-depth analysis of the above problems, this invention proposes an equipment-level load forecasting method and system based on dynamic decoupling of load components to solve the key technical challenges in equipment-level load forecasting. Summary of the Invention

[0015] In view of the above-mentioned problems, the present invention is proposed.

[0016] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a device-level load forecasting method based on dynamic decoupling of load components, comprising: Collect historical load data from the equipment and perform preprocessing; The preprocessed load data is dynamically decomposed by a load component decoupling network to obtain four components and their weights: residential load, commercial load, industrial load and other loads. The load component decoupling network includes a time domain path, a frequency domain path, a feature fusion layer and a soft decomposition layer. Based on a multi-expert model, targeted modeling and prediction are performed for each load component. The multi-expert model includes a group of expert models designed for the characteristics of different load components. The final load prediction result is obtained by adjusting the weight contribution of each expert model through a dynamic gating fusion network, which includes a feature integration layer, a gating calculation layer, and a prediction fusion layer.

[0017] As a preferred embodiment of the equipment-level load forecasting method based on dynamic decoupling of load components described in this invention, the load component decoupling network adopts a time-frequency domain dual-path structure, including: The temporal path, used to capture the temporal dependencies of the load, includes the input layer, feature extraction layer, and temporal coding layer; The frequency domain path is used to extract the periodic features of the load, including a spectrum transform layer, a frequency domain feature extraction layer, and a frequency domain coding layer; Feature fusion layer, used to fuse time-domain and frequency-domain features; The soft decomposition layer is used to generate the weights of each load component.

[0018] As a preferred embodiment of the equipment-level load forecasting method based on dynamic decoupling of load components described in this invention, the soft decomposition layer calculates the weight of each load component using the following formula: in, Indicates the first The weight of each load component, and Representing time-domain and frequency-domain characteristics respectively, Indicates the first Mapping function for each load component This indicates the total number of load components.

[0019] As a preferred embodiment of the equipment-level load forecasting method based on dynamic decoupling of load components described in this invention, the multi-expert model includes: The resident load expert model includes a seasonal-trend decomposition structure, a temperature sensitivity modeling module, a lifestyle habit periodicity identification module, and a time factor encoder; The business load expert model includes dual time-series coding for weekdays and non-weekdays, a business hours feature extractor, a business activity pattern recognizer, and a holiday effect compensation module. The industrial load expert model includes an industry indicator integrator, an economic factor analysis module, a regular production load identifier, and a temporary electricity demand forecasting module. A general load expert model, including an adaptive structure, a multi-feature fusion module, an anomaly pattern recognizer, and a general prediction engine.

[0020] As a preferred embodiment of the equipment-level load forecasting method based on dynamic decoupling of load components described in this invention, the dynamic gating fusion network includes: The feature integration layer is used to integrate static and dynamic features; The gated computation layer is used to calculate the weights of each expert model; The prediction fusion layer is used to fuse the prediction results of various experts based on gating weights.

[0021] As a preferred embodiment of the equipment-level load forecasting method based on dynamic decoupling of load components described in this invention, the gated computation layer calculates the weights of each expert model using the following formula: in, Indicates the first The weights of each expert model, and These represent static features and dynamic features, respectively. Indicates the first The gating function of an expert model. This represents the total number of expert models.

[0022] As a preferred embodiment of the equipment-level load forecasting method based on dynamic decoupling of load components described in this invention, the prediction fusion layer calculates the final prediction result using the following formula: in, This indicates the final prediction result. Indicates the first The prediction results of the expert model Indicates the first The gating weights of an expert model Indicates calibration bias. This represents the total number of expert models.

[0023] A device-level load forecasting system based on dynamic decoupling of load components includes: a preprocessing module for collecting and preprocessing historical load data of equipment; a load decomposition module for dynamically decomposing the preprocessed load data through a load component decoupling network; a load modeling module for modeling and forecasting each load component based on a multi-expert model; and a model fusion module for adjusting the weight contributions of each expert model through a dynamic gating fusion network to obtain the final forecast result.

[0024] A computer device includes: a memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0025] A computer-readable storage medium having a computer program stored thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of the method described in any one of the present invention.

[0026] The beneficial effects of this invention are as follows: First, this invention breaks through the limitations of traditional "one-size-fits-all" prediction methods and proposes a prediction framework based on dynamic decoupling of load components, which significantly improves the prediction accuracy of mixed load equipment. Experimental results show that in prediction scenarios involving equipment with multiple load components, the average prediction error of the method of this invention is reduced by 25%-40% compared with existing single-model methods, especially when the proportion of load components changes rapidly, the improvement in prediction accuracy is even more significant.

[0027] Secondly, this invention provides a deep understanding of the internal structure of mixed loads. Through a load component decoupling network, the system can identify and decompose different types of load components, which not only improves prediction accuracy but also provides more valuable load composition information for power grid planning and demand-side management. This "transparent" approach enhances the interpretability of the model and improves the credibility of the prediction results.

[0028] Third, this invention effectively solves the cold start problem for new equipment. Through the knowledge transfer mechanism of the load component memory, newly commissioned equipment or equipment with sparse data can acquire useful load knowledge from existing equipment, quickly establishing a high-performance prediction model. Experiments show that, with only a small amount of historical data, the method of this invention can control the prediction error within an acceptable range, significantly outperforming traditional methods.

[0029] Fourth, this invention possesses high adaptability to load fluctuations. Through a dynamic gated fusion network, the system can adjust the weight contributions of each expert model in real time according to environmental factors, quickly responding to load pattern changes caused by external conditions. This adaptive fusion mechanism enables the model to maintain high prediction accuracy when facing load fluctuations caused by sudden events, seasonal transitions, etc.

[0030] Fifth, this invention constructs a predictive framework that is capable of sustainable evolution. Through an expert knowledge distillation mechanism and online learning capabilities, the system can continuously accumulate and optimize load knowledge to adapt to the long-term evolution of load characteristics. This continuous evolution capability enables the system to maintain long-term effectiveness when facing emerging load types and changes in electricity consumption behavior.

[0031] In summary, the equipment-level load forecasting method and system based on dynamic decoupling of load components proposed in this invention effectively solves key technical challenges in equipment-level load forecasting by innovatively combining load decoupling, expert modeling, and dynamic fusion, achieving a comprehensive improvement in forecasting accuracy, adaptability, and interpretability. This technical solution is particularly suitable for equipment forecasting scenarios with complex and variable load components, and has significant practical value for improving the safe and economical operation of power systems. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 The first embodiment of the present invention provides an overall flowchart of a device-level load prediction method based on dynamic decoupling of load components.

[0034] Figure 2 This is a schematic diagram of the load component decoupling network of an equipment-level load prediction method based on dynamic decoupling of load components, provided in the first embodiment of the present invention.

[0035] Figure 3This is an architecture diagram of a multi-expert model system for an equipment-level load forecasting method based on dynamic decoupling of load components, provided in the first embodiment of the present invention.

[0036] Figure 4 This is a schematic diagram of the structure of a dynamic gating fusion network for a device-level load prediction method based on dynamic decoupling of load components, provided in the first embodiment of the present invention.

[0037] Figure 5 The diagram illustrates the working principle of a load component memory library for an equipment-level load prediction method based on dynamic decoupling of load components, as provided in the first embodiment of the present invention. Detailed Implementation

[0038] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0039] Example 1, referring to Figures 1-5 As an embodiment of the present invention, a device-level load forecasting method based on dynamic decoupling of load components is provided, such as... Figure 1 As shown, it includes: S1: Load Data Acquisition and Preprocessing. This invention first acquires historical load data and related characteristic data from the equipment, and performs necessary preprocessing to lay the foundation for subsequent load component decoupling and prediction. For example... Figure 1 As shown, data acquisition is the starting point of the entire methodology.

[0040] S1.1: Obtain historical load data for the equipment, including timestamps and corresponding load values. The historical data period is typically the past year, with a sampling frequency of 15 minutes or 1 hour.

[0041] S1.2: Collect environmental characteristic data, including meteorological data (temperature, humidity, wind speed, etc.), time characteristics (time, date, season, holidays, etc.) and equipment characteristics (capacity, type, geographical location, etc.).

[0042] S1.3: Data Cleaning and Anomaly Handling. Outliers are detected and marked using the interquartile range (IMR) method. ,in and These represent the first and third quartiles of the data, respectively. Outliers are imputed or replaced based on the time-series context information.

[0043] S1.4: Data Standardization. Perform Min-Max normalization on the load data: ,in This is the original load value. and These represent the minimum and maximum values ​​in the dataset, respectively. A similar method is used to normalize continuous feature data such as temperature.

[0044] S1.5: Time Feature Encoding. This expands the timestamp into a multi-dimensional time feature, including hourly period encoding. Daily cycle encoding and seasonal cycle coding ,in , and These represent the hour, day of the week, and month, respectively.

[0045] S2: Construction and training of the load component decoupling network. Construction Figure 2 The time-frequency domain dual-path load component decoupling network shown dynamically decomposes the equipment load into four categories: residential, commercial, industrial, and others.

[0046] S2.1: Constructing the temporal path structure. The temporal path employs a multi-scale temporal convolutional network (MS-TCN), containing three one-dimensional convolutional layers with different kernel sizes of 3, 7, and 15, corresponding to short-term, medium-term, and long-term temporal dependencies. Each convolutional layer outputs 64 feature channels, which are then activated by the LeakyReLU function. The formula for extracting time-domain features is: ,in To input timing data, For the first One convolutional kernel.

[0047] S2.2: Constructing the frequency domain path structure. The frequency domain path first decomposes the load signal into different frequency components using Discrete Wavelet Transform (DWT): ,in and These are the approximation coefficients and the detail coefficients, respectively. Then, frequency domain convolution is applied to these coefficients: ,in It is a frequency domain convolution kernel.

[0048] S2.3: Implement the feature fusion layer. The feature fusion layer uses a multi-head attention mechanism to process temporal features. and frequency domain features To achieve interactive integration: ,in This indicates a bullish attention-based operation. , and These represent the query, key, and value matrix, respectively.

[0049] S2.4: Design the soft decomposition layer. The soft decomposition layer maps the fused features to the load component space using a multilayer perceptron (MLP), and generates the weight distribution of each load component through a softmax function: ,in Indicates the first The mapping function of each load component is implemented as a two-layer fully connected network.

[0050] S2.5: Load Component Decoupling Network Training. Due to the lack of direct load component labels, a semi-supervised learning method is adopted. First, a small number of labeled samples are constructed as seed data based on expert knowledge and typical load patterns; then, a contrastive learning strategy is combined to enhance the model's ability to distinguish different load components; finally, the sum of the decoupled components is ensured to be equal to the original load by reconstructing the loss function. ,in For the original load, For the decomposed first Each load component.

[0051] S3: Construction of multi-expert model systems. For example... Figure 3 As shown, specialized and optimized expert models are provided for different types of load components to achieve accurate modeling.

[0052] S3.1: Construction of the Resident Load Expert Model. Based on a structure combining STL (Seasonal-Trend Decomposition using Loess) and LSTM networks, the steps include: First, decomposing the resident load into seasonal components using STL. Trend components and residual components : Then, a periodic neural network is used to model the seasonal component, an LSTM network is used to model the trend component, and a GRU network is used to model the residual component. Finally, the three prediction results are added together to obtain the final prediction. The model also integrates a temperature-sensitive module, using a nonlinear function. Modeling the asymmetric effect of temperature on load, among which For temperature, As the reference temperature, and These are learnable parameters.

[0053] S3.2: Construction of the Business Load Expert Model. A dual-stream neural network structure is employed, comprising two parallel processing paths: a time-stream and a feature-stream. The time-stream handles the time encoding of weekdays and non-weekdays. ,in Time-encoded features; feature stream processing for business activity features: ,in Features of business activities are then incorporated from the two streams using an attention mechanism. Finally, a business load forecast is generated through a fully connected layer: The model also includes a dedicated holiday effect compensation module, which captures the unique impact of different holidays on business load through learnable holiday embedding vectors.

[0054] S3.3: Construction of the Industrial Load Expert Model. A structure combining Graph Convolutional Networks (GCNs) and time-series forecasting is adopted. First, an industry linkage graph is constructed. , where nodes To indicate different industry types, edge This represents the strength of inter-industry linkages. Then, graph convolution operations are used to extract industry linkage features. ,in To add self-connected adjacency matrices, for The degree matrix, For the first Layer node characteristics, This is a learnable weight matrix. Finally, the temporal evolution of industrial load is modeled using the time series forecasting module: ,in Represents a temporal convolutional network. The model provides input characteristics for industrial load. It also integrates an economic factor analysis module, using an adaptive gating mechanism to adjust for the impact of macroeconomic indicators on industrial load.

[0055] S3.4: Construction of a General Load Expert Model. Based on the meta-learning framework MAML (Model-Agnostic Meta-Learning), it can quickly adapt to new load patterns. The model architecture consists of two parts: a feature extraction backbone and a prediction head. During the meta-training phase, task sets are constructed for different load patterns. Each task contains a support set and query set The model parameters are updated through a two-layer optimization process: the inner loop performs gradient updates on the support set for each task. The outer loop performs meta-optimization on the query set of all tasks. ,in and These are the learning rates for the inner and outer loops, respectively.

[0056] S4: Construction of Dynamic Gated Fusion Network. Based on... Figure 4 The structure shown is used to construct a dynamic gating fusion network to achieve adaptive fusion of prediction results from various expert models.

[0057] S4.1: Feature Integration Layer Design. The feature integration layer consists of two parts: a static feature encoder and a dynamic feature encoder. The static feature encoder processes inherent characteristics of the device (such as location, type, and capacity): ,in The static feature vector is used; the dynamic feature encoder processes time-varying environmental factors (such as time, weather, and external events). ,in This is a dynamic feature sequence. Then, the feature fusion module integrates these two types of features: ,in Gating mechanism adopted: , As a gated vector, through calculate.

[0058] S4.2: Gated Computation Layer Design. The gated computation layer employs a self-attention-enhanced multilayer perceptron to calculate the weights of each expert model. First, the feature representation is enhanced through a self-attention mechanism: in, , The dimension is the key. Then, the raw scores of each expert model are calculated using a multilayer perceptron: Finally, the normalized weight coefficients are generated using the Softmax function: ,in This represents the total number of expert models.

[0059] S4.3: Prediction Fusion Layer Design. The prediction fusion layer fuses the prediction results from various experts based on gating weights, and performs a final adjustment to the fusion result through a calibration network. The fusion calculation formula is as follows: ,in For the first The prediction results of the expert model For the corresponding gating weights, For bias calibration. The calibration network uses a residual connection structure: This allows for fine-tuning while preserving the original fusion information.

[0060] S4.4: Gated Network Training. The training objective of the gated network is to minimize the error between the final prediction and the actual load. The loss function uses a combination of mean squared error and mean absolute error: ,in and The weights are set to 0.6 and 0.4 respectively. The training process uses the backpropagation algorithm to update the network parameters, with the initial learning rate set to 0.001 and dynamically adjusted using a cosine annealing strategy.

[0061] S5: Construction of the load component memory. For example... Figure 5 As shown, a load component memory is established to provide transfer learning capabilities for data-sparse devices.

[0062] S5.1: Memory organization structure design. Based on the Prototypical Network organization structure, indexes are established according to device type and load component type. Each device type maintains a set of prototype vectors: ,in For device type and load components prototype vector, For the corresponding support set, This is the feature extraction function. The prototype vector captures the typical feature patterns of this type of device and load components.

[0063] S5.2: Similarity retrieval mechanism implementation. The similarity calculation between the query device and samples in the memory bank uses an enhanced cosine similarity formula: ,in and These are the feature vectors of the query device and the memory sample, respectively. As a measure of contextual difference, and These are the context feature vectors for the query device and the memory sample, respectively. and These are adjustable parameters, set to 0.5 and 2.0 respectively. The retrieval process returns the Top-K samples with the highest similarity.

[0064] S5.3: Knowledge Transfer Module Design. Knowledge transfer employs a progressive learning method, comprising three stages: Initialization stage, which initializes the target device model based on the load component patterns of similar samples in the memory bank. ,in For the first Model parameters for similar samples, Weights are based on similarity; in the fine-tuning phase, model parameters are adjusted using a small amount of target device data. ,in For the target device's dataset, The learning rate; during the online learning phase, the model parameters are continuously updated as target device data accumulates. ,in For time steps Model parameters, For the new data points, This is the dynamic learning rate.

[0065] S5.4: Memory Maintenance and Updates. The memory is kept up-to-date through a periodic update mechanism. Whenever a new device is processed, its load composition pattern is added to the memory; for existing devices, if the load pattern changes significantly (similarity below a threshold), the memory is updated. If a sample is found to be a frequently used sample, its record in the memory bank is updated. To prevent the memory bank from growing indefinitely, a eviction strategy based on usage frequency and representativeness is adopted to retain the most representative and frequently used samples.

[0066] S6: Implementation of Expert Knowledge Distillation Mechanism. Implementing an expert knowledge distillation mechanism allows expert models to learn from each other's experiences, improving overall predictive capabilities.

[0067] S6.1: Knowledge Extraction and Soft Label Generation. Prediction results are collected from various expert models, and soft labels are generated using a temperature-scaled softmax function. ,in For the first The output logits of each expert The temperature parameter (set to 2.0) controls the degree of "softening" of the soft tags. Integrated soft tags are generated using a weighted average. ,in For the first The reliability weights of each expert are calculated based on their historical prediction accuracy.

[0068] S6.2: Distillation Loss Calculation and Optimization. The distillation loss is calculated using a weighted combination of the true label loss and the knowledge distillation loss: ,in For cross-entropy loss, Let KL divergence be the KL divergence. This is the balancing parameter (set to 0.7). The expert model learns useful knowledge from other experts by minimizing this loss function.

[0069] S6.3: Implementation of Selective Knowledge Learning. To avoid negative interference, a selective knowledge learning mechanism is introduced. Knowledge transfer is controlled through a gating mechanism: , ,in and Experts and experts The hidden layer representation. To the experts To experts The knowledge gating coefficient, , and This is a learnable parameter. When the gating coefficient is approximately 0, it indicates that the transmission of the corresponding knowledge is blocked; when it is close to 1, it indicates that the corresponding knowledge is fully received.

[0070] S7: Implementation of the overall algorithm flow. According to... Figure 1 The process shown implements an equipment-level load forecasting algorithm based on dynamic decoupling of load components.

[0071] S7.1: Data Preprocessing Stage. The raw load data and feature data are cleaned, standardized, and feature-engineered to generate model input data.

[0072] S7.2: Load Component Decoupling Stage. The preprocessed load data is dynamically decomposed through a load component decoupling network to obtain four components: residential load, commercial load, industrial load, and other loads, along with their weight distributions.

[0073] S7.3: Multi-expert modeling stage. Each decoupled load component is input into its corresponding expert model for prediction, yielding prediction results for each component. For new equipment or equipment with sparse data, a load component memory is enabled for knowledge transfer.

[0074] S7.4: Dynamic Fusion Stage. Through a dynamically gated fusion network, the weight contributions of each expert model are dynamically adjusted based on the current environmental conditions to obtain the final load forecast result.

[0075] S7.5: Model Update and Optimization Stage. Based on prediction error feedback, the parameters of each module are updated through online learning; simultaneously, an expert knowledge distillation mechanism promotes knowledge sharing and collaborative optimization among expert models.

[0076] S8: System Implementation Plan. Construct a complete equipment-level load forecasting system based on dynamic decoupling of load components.

[0077] S8.1: Employs a distributed computing architecture, allocating complex computing tasks across multiple computing nodes. Load component decoupling and expert model prediction tasks can be executed in parallel, with the master node responsible for result integration and coordination. The system architecture comprises three parts: a data layer, a computing layer, and an application layer, supporting horizontal scaling to adapt to ever-increasing computing demands.

[0078] S8.2: Designed with multi-granularity forecasting capabilities, supporting short-term (hourly), medium-term (daily), and long-term (weekly / monthly) load forecasts. Different feature combinations and model configurations are used for different forecast periods. Short-term forecasts focus more on immediate environmental factors, while long-term forecasts focus more on macroeconomic trends and seasonal patterns.

[0079] S8.3: Implements a visual interface to display the decoupling results and prediction process of load components. The interface includes functional modules such as load component proportion analysis, prediction result display, error analysis, and key influencing factor analysis, improving system transparency and interpretability.

[0080] Example 2, an embodiment of the present invention, provides an equipment-level load forecasting system based on dynamic decoupling of load components, including: a preprocessing module, a load decomposition module, a load modeling module, and a model fusion module; The preprocessing module is used to collect and preprocess historical load data of the equipment; the load decomposition module is used to dynamically decompose the preprocessed load data through a load component decoupling network; the load modeling module is used to model and predict each load component based on a multi-expert model; and the model fusion module is used to adjust the weight contribution of each expert model through a dynamic gating fusion network to obtain the final prediction result.

[0081] Example 3, an embodiment of the present invention, differs from the previous two embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, 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 steps 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 USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0082] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0083] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0084] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0085] Example 4: To verify the device-level load forecasting method based on dynamic decoupling of load components proposed in this invention, the experiment was conducted at the distribution network data center of a provincial power company. Five different types of substations were selected as test objects, including a substation in a comprehensive commercial area, a substation in a residential area, a substation in an industrial park, a substation in a school park, and a substation in a mixed commercial and residential area. The hardware environment used an Intel Xeon Gold 6248R processor (3.0GHz, 48 cores), 384GB of memory, and an NVIDIA Tesla A100 GPU for acceleration. The software environment used Python 3.9 and the PyTorch 2.0 framework. The experiment used historical load data from January 2023 to December 2023, with a time granularity of 15 minutes, and collected meteorological data, holiday information, and relevant economic indicators for the same period.

[0086] Traditional single neural network prediction method (LSTM-FCN): This method employs a hybrid model combining long short-term memory networks and fully convolutional networks to directly model and predict the overall load, without considering differences in load components. The model structure consists of two LSTM layers (each with 128 neurons) and three convolutional layers (with kernel sizes of 3, 5, and 7), finally outputting the prediction result through a fully connected layer. This method represents the current mainstream load prediction technology approach.

[0087] Load Classification Forecasting Method (Load-Type-Net): First, loads are classified into three categories—residential, commercial, and industrial—using manual rules. Then, an independent forecasting model is trained for each category. Residential loads use a seasonal ARIMA model, commercial loads use a gradient boosting tree model, and industrial loads use a deep neural network model. Finally, the forecast results from each model are summed with fixed weights (determined based on historical average proportions) to obtain the final forecast result. This method represents a technical approach that considers load type but uses static classification.

[0088] First, data preprocessing is performed, including a comprehensive cleaning of the collected historical load data, outlier removal using a moving median filter, and missing data points filled in using linear interpolation. Then, multidimensional time features are extracted, including hour, date, day of the week, month, and holiday indicators. Simultaneously, a meteorological feature vector is constructed, integrating environmental factors such as temperature, humidity, wind speed, and light intensity. All feature data undergoes Min-Max standardization to ensure that feature values ​​for each dimension are mapped to the [0,1] interval, eliminating the influence of unit dimensions.

[0089] After data preparation, the entire dataset is divided into training, validation, and test sets in an 8:1:1 ratio according to time sequence. To implement the scheme of this invention, the load component decoupling network is first trained, employing a dual-path architecture to achieve automatic decomposition of mixed loads. The time-domain path uses a three-layer multi-scale temporal convolutional structure of 64-128-256, with receptive fields of 3, 9, and 27, respectively, effectively capturing features at different time scales. The frequency-domain path extracts multi-scale frequency-domain features through a 5-layer discrete wavelet transform, which are then effectively represented by a frequency-domain feature extraction network. The features from the two paths are fused through a multi-head attention mechanism (8 heads), and then the soft decomposition layer generates the weight distribution of each load component.

[0090] Next, specialized optimized expert models were trained for each load component type. The residential load expert model employs an architecture combining STL decomposition and a bidirectional LSTM network. Seasonal, trend, and residual components are modeled using LSTM networks with 32-64-128 neurons respectively, and the final results are integrated through an adaptive fusion network. The commercial load expert model uses a two-stream neural network structure. The time stream is processed by a GRU network with 128-256 neurons, while the feature stream extracts features through a three-layer residual network. The two streams of features are organically fused through a cross-attention mechanism. The industrial load expert model utilizes a graph convolutional network constructed from a supply chain relationship graph, combined with a TCN network with 64-128 nodes to capture periodic production characteristics. The general load expert model is implemented based on the MAML framework and achieves rapid adaptation through five gradient updates.

[0091] Subsequently, a dynamic gating fusion network is trained to learn the optimal integration method between static features (such as device location, type, and capacity) and dynamic features (such as time, weather, and external events). The feature integration layer adopts a multilayer perceptron with 128-256-512 nodes, and the gating computation layer introduces a 16-head self-attention mechanism to enhance key features. The prediction fusion layer achieves final weight allocation and bias calibration through a three-layer fully connected network.

[0092] A load component memory is constructed in parallel, containing typical load patterns of various equipment over the past year. It is indexed and organized by equipment type, storing 100 prototype vectors for each equipment type. The memory employs a similarity retrieval mechanism, using an improved cosine similarity formula for matching, providing knowledge transfer for new equipment. Finally, an expert knowledge distillation process is implemented, with a temperature parameter set to 2.0 and a balancing parameter... Setting it to 0.7 allows each expert model to draw on the strengths and experiences of other experts while maintaining its specialization.

[0093] After training is completed, the proposed solution and two comparative solutions are comprehensively evaluated and tested on the test set, covering predictions at different time granularities (15 minutes, 1 hour, 1 day), including different load scenarios such as weekdays, weekends, holidays and extreme weather, and targeted tests of load component mutation scenarios, such as large-scale promotional activities in shopping malls, temporary production changes in factories and sudden events.

[0094] This experiment employs a multi-dimensional evaluation system to comprehensively measure the performance of each prediction scheme. Prediction accuracy is evaluated using the Mean Absolute Percentage Error (MAPE) as the core indicator, calculated using the following formula: ,in This represents the actual load value. This represents the predicted load value. This represents the number of predicted samples. The root mean square error (RMSE) is also introduced to evaluate the absolute error; the calculation formula is as follows: The coefficient of determination (R²) assesses the model's ability to explain load changes, and is calculated using the following formula: ,in This represents the average value of the actual load.

[0095] For robustness assessment, three specialized test scenarios were designed: load mutation adaptability test, which simulates scenarios such as commercial promotions and sudden changes in industrial production, and records the time required for the prediction error to recover to normal levels; abnormal weather adaptability test, which selects the prediction accuracy under extreme meteorological conditions such as extreme high temperature, cold wave, and rainstorm; and special event adaptability test, which targets the prediction accuracy during holidays such as Spring Festival and National Day, as well as special events such as large-scale sports events.

[0096] Computational efficiency evaluation included training time measurement (time required to complete the entire training process, in hours), prediction speed measurement (time required to generate one day's prediction results, in seconds), and resource consumption monitoring (memory usage and GPU utilization during training and prediction). During testing, each metric was measured 10 times, the average was taken, and the standard deviation was calculated to ensure statistical significance of the results.

[0097] The generalizability assessment focuses on the model's generalization ability and adaptability, including cross-equipment applicability testing (prediction accuracy on different types of substations), cold start capability testing (prediction accuracy during the initial commissioning of new equipment, with data volumes of 10%, 20%, and 50% of the standard amount), and interpretability assessment (the degree of agreement between the load component decoupling results and the actual load composition, verified through manual annotation). To ensure the comprehensiveness of the assessment, equipment with different load characteristics (high residential proportion, high commercial proportion, high industrial proportion, and balanced mixed proportion) are tested separately, and the results are weighted and averaged, with the weights determined according to the actual distribution proportion of various types of equipment in the power grid.

[0098] Table 1. Experimental Results

[0099] As shown in Table 1, the proposed solution achieves a prediction error of 2.17% on the overall MAPE index, representing improvements of 62.8% and 48.5% compared to comparative solutions 1 and 2, respectively. This significant improvement stems from the load component decoupling network's ability to accurately identify the inherent structural features of mixed loads, providing precise load components for subsequent expert models. The time-frequency dual-path design enables the decoupling network to simultaneously capture the temporal dependence and periodicity of the load, greatly improving the accuracy of component identification, as evidenced by the 91.8% load component identification accuracy.

[0100] In tests conducted at various types of substations, the proposed solution demonstrated comprehensive performance advantages. Particularly in mixed-area scenarios, the MAPE of this solution was only 2.34%, representing improvements of 73.8% and 65.7% compared to Comparative Solution 1 and Comparative Solution 2, respectively. This fully demonstrates the superior capability of this solution in handling complex load compositions. Prediction errors for residential, commercial, and industrial areas were 2.08%, 2.36%, and 2.51%, respectively, all showing significant improvements. This result validates the effectiveness of the multi-expert model system; each expert model, through optimized design for specific load components, can more accurately capture the characteristics and variation patterns of that type of load.

[0101] In load mutation adaptability tests, the proposed solution demonstrated outstanding performance. Under scenarios simulating rapid changes in load components, such as commercial promotions and industrial production fluctuations, the proposed solution restored normal prediction accuracy in just 1.8 hours, representing reductions of 76.0% and 65.4% compared to comparative solutions 1 and 2, respectively. This is attributed to the dynamic gating fusion network's ability to adjust the weight contributions of each expert model in real time based on input conditions, rapidly adapting to changes in load patterns. The self-attention mechanism in the gating network effectively enhances the emphasis on key features, improving the accuracy of weight allocation.

[0102] The seasonal variation RMSE test reflects the model's adaptability to load pattern changes caused by seasonal transitions. The RMSE value of the proposed solution is 103.4 kW, which is 68.5% and 57.2% lower than that of comparative solutions 1 and 2, respectively. This indicates that the load component decoupling method of the present invention can effectively identify the variation characteristics of each load component under different seasons, and the expert model system can specifically model these changes. In particular, the STL decomposition structure in the residential load expert model shows strong modeling ability for seasonal variations.

[0103] Load extreme value prediction is crucial for the safe operation of power systems. The proposed solution achieves deviations of 2.12% and 2.04% in peak and valley load prediction, respectively, representing a significant improvement over comparative solutions. This is of great value for grid dispatching and capacity planning, effectively avoiding resource waste and system risks. The prediction accuracy for special events reaches 92.7%, superior to the 78.4% and 83.6% of the comparative solutions, demonstrating the invention's stronger adaptability to unconventional load scenarios.

[0104] The cold start capability test reflects the model's ability to handle new equipment with sparse data. The proposed solution, with only a small amount of historical data, predicts a MAPE of only 3.62%, a reduction of 70.7% and 63.2% compared to Comparative Solution 1 and Comparative Solution 2, respectively. This superior performance stems from the knowledge transfer mechanism of the load component memory, which can extract similar load patterns from existing equipment and adapt them to new equipment, quickly building a high-quality prediction model. The proposed solution achieves high prediction accuracy without requiring a large amount of historical data, significantly lowering the implementation threshold for predicting new equipment.

[0105] In terms of computational efficiency, the prediction time of the proposed solution is 0.63 seconds / day, slightly higher than that of comparative solution 1 (0.42 seconds / day), but lower than that of comparative solution 2 (0.75 seconds / day). Considering the significant improvement in prediction accuracy, this computational cost is entirely reasonable. The number of model parameters is 6.8 million, falling between that of comparative solution 1 (4.2 million) and comparative solution 2 (9.7 million), indicating that the proposed solution achieves a substantial performance improvement while maintaining a reasonable level of model complexity.

[0106] In cross-device applicability testing, the average MAPE of the proposed solution was 2.43%, which is 66.3% and 57.8% higher than that of comparative solution 1 and comparative solution 2, respectively. This demonstrates that the proposed solution has strong generalization ability and can adapt to different types and characteristics of equipment. This generalization ability stems from the deep understanding of equipment characteristics through the load component decoupling mechanism, as well as the knowledge sharing and transfer learning capabilities promoted by the expert knowledge distillation mechanism.

[0107] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0108] Example 5 provides an extended implementation scheme for a dynamic load component decoupling system in a new energy integration scenario. With the continuous increase in installed capacity of new energy sources such as wind power and photovoltaics, the load characteristics of the power system have changed significantly, making traditional load forecasting methods difficult to adapt to this new mixed load structure. To address this issue, this example extends the basic technical solution to new energy consumption scenarios, enhancing the system's forecasting capability under conditions of high renewable energy penetration.

[0109] In this embodiment, the basic load component decoupling network is first extended to enhance its ability to identify renewable energy loads. The modified load component decoupling network can identify not only traditional residential, commercial, industrial, and other loads, but also two newly added components: wind power consumption load and photovoltaic consumption load. A new energy feature extraction module is added to the time-domain path, employing a deep residual network structure to capture the temporal characteristics of new energy loads; a wavelet packet transform layer is added to the frequency-domain path to extract the frequency domain features of new energy loads with finer frequency resolution. The decomposition formula for the soft decomposition layer is correspondingly extended as follows: in, Indicates the first The weight of each load component, and Representing time-domain and frequency-domain characteristics respectively, the newly added Indicates characteristics related to new energy. Indicates the first Mapping function for each load component This represents the total number of expanded load components (the original N types plus wind and solar power). To improve the accuracy of renewable energy load identification, a mapping function is used. By employing a depthwise separable convolutional structure, the expressive power of the model is improved while maintaining computational efficiency.

[0110] For the newly added wind power and photovoltaic power consumption loads, this embodiment designs two dedicated expert models. The wind power consumption load expert model adopts a meteorological-physical fusion architecture, comprising three parts: a weather factor analysis module, a wind power conversion module, and a load response module. The weather factor analysis module receives meteorological data such as wind speed, wind direction, air pressure, and temperature, and generates weather feature vectors through a multilayer perceptron network. The wind power conversion module establishes a nonlinear mapping relationship between wind speed and power generation based on the wind turbine characteristic curve and the regional wind power installed capacity, generating a power feature vector. The load response module analyzes the grid's response characteristics to wind power absorption and generates response feature vectors. The final wind power load forecast is calculated using the following formula: in, This represents the predicted wind power absorption load. Represents the neural network mapping function. This represents the concatenation of eigenvectors. For calibration coefficients, This is the wind power capacity factor, used to adapt to wind power scenarios of different scales.

[0111] The photovoltaic load expert model employs an irradiance-driven dual-attention structure, comprising an irradiance prediction module, a photovoltaic conversion efficiency module, and a grid integration module. The irradiance prediction module combines astronomical calculations and meteorological conditions to predict solar irradiance for future periods. The photovoltaic conversion efficiency module considers factors such as temperature, module type, and degradation to calculate the actual power generation efficiency. The grid absorption module analyzes the grid's ability to absorb photovoltaic power and generates an absorption coefficient. The outputs of each module are weighted using both spatial and temporal attention mechanisms to obtain the final photovoltaic load forecast. in, This represents the predicted value of photovoltaic load. and These are spatial attention weights and temporal attention weights, respectively. For photovoltaic installed capacity, For the predicted duration.

[0112] To adapt to the drastic fluctuations in load composition ratios under high renewable energy penetration scenarios, this embodiment enhances the dynamic gating fusion network. First, a meteorological factor encoder and a renewable energy output encoder are added to the feature integration layer to encode weather changes and renewable energy output, respectively. Then, a multi-scale time gating mechanism is introduced into the gating calculation layer, capable of capturing load composition variation patterns at different time scales from minutes to days. The improved gating calculation formula is as follows: in, Indicates the first On the first time scale The weights of each expert model, and These are static and dynamic features, respectively, and are newly added. As a meteorological feature, Characteristics of contributing to new energy sources For the first On the first time scale The gating function of an expert model. The total number of expert models, For the first Importance coefficients for each time scale For the final gating weight, This represents the total number of time scales.

[0113] The prediction fusion layer has also been expanded accordingly, adding a new energy output correction mechanism that can dynamically correct the prediction results based on real-time new energy output. in, This indicates the final prediction result. Indicates the first The prediction results of the expert model Indicates the first The gating weights of an expert model and They represent the first The actual and projected output of new energy sources This represents the corresponding correction factor. Indicates calibration bias. This represents the total number of expert models.

[0114] To address the cold start issue in renewable energy scenarios, this embodiment expands the load component memory. The memory adds a renewable energy load pattern library, organizing renewable energy load patterns by region type, installed capacity structure, and absorption characteristics. The similarity retrieval module adds a meteorological condition matching mechanism, calculating the similarity of equipment in renewable energy scenarios using the following formula: in, This indicates the similarity between devices X and Y in a new energy scenario. This represents the basic similarity calculation formula. Indicates the similarity of meteorological conditions. and Let X and Y represent the meteorological feature vectors of devices X and Y, respectively. Indicates the similarity of installed capacity structure. and Let X and Y represent the feature vectors of installed capacity, respectively. and This indicates an adjustable parameter.

[0115] The knowledge transfer module has also been enhanced, adding a new energy penetration rate adaptation layer, which can automatically adjust transfer parameters based on the new energy penetration rate level of the target equipment. in, and These represent the model parameters of the target device and the source device, respectively. Indicates the learning rate. This represents a matrix showing the differences in new energy penetration rates. This represents the gradient of the source device model. This adaptive migration mechanism can effectively handle the forecasting needs of different new energy penetration rates.

[0116] This embodiment also extends the expert knowledge distillation mechanism by adding a new energy expert knowledge extraction module. This module specifically extracts key knowledge from wind power and photovoltaic expert models and passes it to other expert models to improve their predictive capabilities in new energy scenarios. The extended distillation loss formula is as follows: in, This indicates distillation losses in new energy scenarios. Represents cross-entropy loss, Denotes KL divergence, This indicates the integration of soft tags. Soft labels indicating new energy expert models Indicates the first A soft label for an expert and Represents the balance parameters. Indicates the true label, Indicates the first The predictions of several experts.

[0117] To meet the high-frequency requirements of new energy forecasting, this embodiment implements multi-time-granularity forecasting capabilities, including ultra-short-term (5-minute), short-term (hourly), medium-term (daily), and long-term (weekly / monthly) forecasts. Each time granularity employs a cascaded forecasting structure, with high-frequency forecast results serving as input features for low-frequency forecasts, and is integrated using the following formula: in, Indicates the first Prediction results at each time granularity This represents the corresponding prediction function. Represents the original input features. This indicates prediction results at a higher frequency time granularity.

[0118] This embodiment also introduces an uncertainty quantification mechanism to provide a reliable confidence interval for each prediction result. By integrating Monte Carlo Dropout technology, the system can estimate the distribution characteristics of the predictions: in, This indicates the final prediction result. Indicates the first Prediction results after random Dropout Indicates the prediction variance. Indicates the model accuracy parameters. This represents the number of Monte Carlo samplings. Based on this, we can construct... Confidence interval: in, For standard normal distribution Quantiles.

[0119] Finally, this embodiment establishes a new energy-load joint optimization mechanism to guide new energy consumption strategies through forecast results. Based on load forecasts and new energy output forecasts, the system automatically generates optimized new energy dispatch suggestions to improve the new energy consumption rate and reduce wind and solar curtailment rates, providing decision support for grid dispatch.

[0120] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A device-level load forecasting method based on dynamic decoupling of load components, characterized in that, include: Collect historical load data from the equipment and perform preprocessing; The preprocessed load data is dynamically decomposed by a load component decoupling network to obtain four components and their weights: residential load, commercial load, industrial load and other loads. The load component decoupling network includes a time domain path, a frequency domain path, a feature fusion layer and a soft decomposition layer. Based on a multi-expert model, targeted modeling and prediction are performed for each load component. The multi-expert model includes a group of expert models designed for the characteristics of different load components. The final load prediction result is obtained by adjusting the weight contribution of each expert model through a dynamic gating fusion network, which includes a feature integration layer, a gating calculation layer, and a prediction fusion layer.

2. The equipment-level load forecasting method based on dynamic decoupling of load components as described in claim 1, characterized in that: The load component decoupling network adopts a time-frequency domain dual-path structure, including: Temporal path, frequency path, feature fusion layer, soft decomposition layer.

3. The equipment-level load forecasting method based on dynamic decoupling of load components as described in claim 2, characterized in that: The soft decomposition layer calculates the weights of each load component using a formula, expressed as follows: in, Indicates the first The weight of each load component, and Representing time-domain and frequency-domain characteristics respectively, Indicates the first Mapping function for each load component This indicates the total number of load components.

4. The equipment-level load forecasting method based on dynamic decoupling of load components as described in claim 3, characterized in that: The multi-expert model includes: Residential load expert model, commercial load expert model, industrial load expert model, general load expert model.

5. The equipment-level load forecasting method based on dynamic decoupling of load components as described in claim 4, characterized in that: The dynamic gating fusion network includes: a feature integration layer that integrates static and dynamic features; a gating computation layer that calculates the weights of each expert model; and a prediction fusion layer that fuses the prediction results of each expert based on the gating weights.

6. The equipment-level load forecasting method based on dynamic decoupling of load components as described in claim 5, characterized in that: The gated computation layer uses a formula to calculate the weights of each expert model, expressed as: in, Indicates the first The weights of each expert model, and These represent static features and dynamic features, respectively. Indicates the first The gating function of an expert model. This represents the total number of expert models.

7. The equipment-level load forecasting method based on dynamic decoupling of load components as described in claim 6, characterized in that: The prediction fusion layer calculates the final prediction result using a formula, expressed as follows: in, This indicates the final prediction result. Indicates the first The prediction results of the expert model Indicates the first The gating weights of an expert model Indicates calibration bias. This represents the total number of expert models.

8. A system for equipment-level load forecasting based on dynamic decoupling of load components, employing the equipment-level load forecasting method based on dynamic decoupling of load components as described in any one of claims 1 to 7, characterized in that, include: The preprocessing module is used to collect and preprocess historical load data of the equipment; The load decomposition module is used to dynamically decompose the preprocessed load data through a load component decoupling network. The load modeling module is used to model and predict each load component based on a multi-expert model. The model fusion module is used to adjust the weight contributions of each expert model through a dynamic gating fusion network to obtain the final prediction result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the equipment-level load forecasting method based on dynamic decoupling of load components as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the equipment-level load forecasting method based on dynamic decoupling of load components as described in any one of claims 1 to 7.

Images