Intelligent prediction method and system for electrical load, and storage medium

Abstract

The invention relates to the technical field of power consumption load prediction, and discloses a power consumption load intelligent prediction method and system, and a storage medium. The method comprises the following steps: carrying out regional difference quantization on load data of a source domain and a target domain to obtain a regional feature vector; carrying out time sequence coding and attention calculation on the source domain load to obtain source domain load time sequence characteristics and weight distribution; domain invariant load features are obtained through domain confrontation training; performing cross-domain migration in combination with knowledge fusion to obtain target domain load features; and generating a cross-regional load prediction result through a multi-step prediction algorithm after fusing the historical data of the target domain and performing fine tuning. The problem that the prediction precision is reduced due to the feature distribution difference of the source domain and the target domain in cross-regional load prediction is solved.

Description

Technical Field

[0001] This application relates to the field of electricity load forecasting technology, and in particular to an intelligent electricity load forecasting method, system and storage medium. Background Technology

[0002] Existing electricity load forecasting technologies primarily rely on historical data from a single region for modeling, employing methods such as time series analysis and machine learning to predict local electricity load. These methods typically require a large amount of historical load data for model training and can achieve good predictive results in areas with sufficient data.

[0003] However, existing technologies have significant shortcomings when dealing with newly established regions or target regions with scarce historical data: on the one hand, new regions lack sufficient historical data to support model training, resulting in low prediction accuracy; on the other hand, when directly applying models trained in other regions to the target region, the prediction results often deviate significantly due to differences in industrial structure, electricity consumption habits, and climate conditions between different regions. Furthermore, existing methods lack effective cross-regional knowledge transfer mechanisms and cannot fully utilize the prediction experience of data-rich regions.

[0004] Even with simple transfer learning methods, the following progressive technical problems still exist: First, the differences in the distribution of load features between the source and target domains are difficult to eliminate effectively, resulting in regional bias in the transferred features; second, there is a lack of targeted knowledge screening mechanisms, making it impossible to identify which source domain knowledge is truly useful for predicting the target domain; finally, the transferred model lacks integration and adjustment with the actual situation of the target domain, ignoring the local characteristics of the target domain. Summary of the Invention

[0005] This application provides an intelligent load forecasting method, system, and storage medium to solve the problem of decreased forecasting accuracy caused by the difference in feature distribution between the source and target domains in cross-regional load forecasting. It also solves the problem of poor migration effect due to the lack of effective knowledge screening and fusion mechanisms even after eliminating distribution differences. Finally, it solves the problem that newly built or data-scarce areas cannot establish accurate forecasting models due to the lack of sufficient historical data.

[0006] In a first aspect, this application provides an intelligent load forecasting method, which includes: quantifying the regional differences in load data of the source and target domains using a regional load pattern recognition algorithm to obtain a regional feature vector containing load factor, peak-valley difference rate, and load fluctuation coefficient; performing time-series encoding and attention weight calculation on the regional feature vector to obtain source domain load time-series features and load attention weight distribution; performing domain adaptation processing on the source domain load time-series features through domain adversarial training to obtain domain-invariant load features; performing cross-domain migration processing on the domain-invariant load features and load attention weight distribution according to a knowledge fusion algorithm to obtain target domain load features; fusing and fine-tuning the target domain historical load data with the target domain load features to obtain optimized target domain load features; and performing time-series prediction processing on the optimized target domain load features according to a multi-step prediction algorithm to generate cross-regional load forecasting results.

[0007] Optionally, the step of performing regional difference quantification processing on the load data of the source and target domains using a regional load pattern recognition algorithm to obtain a regional feature vector containing load factor, peak-to-valley difference rate, and load fluctuation coefficient includes: Outlier detection and missing value interpolation are performed on the historical electricity load data of the source and target domains to obtain standardized load time-series data. Based on the standardized load time-series data, the average load value and the maximum load value are calculated, and the ratio of the average load value to the maximum load value is calculated to obtain the load rate value. Based on the standardized load time series data, peak load and valley load are extracted. The difference between peak load and valley load is divided by the average load value to calculate the peak-valley difference rate, and the peak-valley difference rate value is obtained. The load standard deviation is calculated based on the standardized load time series data. The load standard deviation is then compared with the average load value to obtain the load fluctuation coefficient. The load rate, peak-valley difference rate, and load fluctuation coefficient are then combined into a vector to obtain the regional feature vector.

[0008] Optionally, the step of performing temporal encoding and attention weight calculation on the regional feature vector to obtain the source domain load temporal features and load attention weight distribution includes: The region feature vectors are segmented in chronological order using a sliding window mechanism to obtain load feature sequence segments of fixed length. The load characteristic sequence segment is subjected to multi-scale time series decomposition processing to extract short-term fluctuation components, daily cycle components and long-term trend components respectively. The three components are then fused to obtain a multi-scale time series representation. Based on the multi-scale temporal representation, a position encoder generates a position encoding vector according to the relative position of each time step. The position encoding vector is then superimposed with the multi-scale temporal representation to obtain the source domain load temporal features. The correlation scores between different time steps of the source domain load time series features are calculated using a self-attention mechanism to obtain the time series correlation matrix; The temporal correlation matrix is ​​normalized so that the sum of the attention weights at each time step is a unit value, thus obtaining the load attention weight distribution.

[0009] Optionally, the step of performing domain adaptation processing on the source domain load time-series features through domain adversarial training to obtain domain-invariant load features includes: The source domain load time-series features are input into the load feature extraction network for deep feature extraction processing, and the target domain regional load data are input into the same load feature extraction network for feature extraction processing to obtain the source domain load deep features and the target domain load deep features. Based on the deep features of the source domain load and the deep features of the target domain load, the load source region is classified by the region discriminator, the region classification loss value is calculated, and the gradient inversion layer is used to backpropagate the gradient of the region classification loss to obtain the inverted gradient signal used to confuse regional differences. The parameters of the load feature extraction network are updated according to the inverted gradient signal, so that the load features generated by the load feature extraction network cannot be identified by the region discriminator as their source region, thus obtaining region-independent load confusion features. The region-independent load confusion features are processed by calculating the inter-regional load pattern distance through cross-regional load distribution difference measurement. Training stops when the inter-regional load pattern distance is less than a preset convergence threshold, thus obtaining the domain-invariant load features.

[0010] Optionally, the step of performing cross-domain transfer processing on the domain-invariant load features and load attention weight distribution according to the knowledge fusion algorithm to obtain the target domain load features includes: The domain-invariant load features are compressed using a knowledge distillation network to obtain compressed common load knowledge. Based on the load attention weight distribution, the common load knowledge is weighted and selected, and knowledge fragments with high relevance to the target domain load pattern are selected according to the attention weight value to obtain target domain related load knowledge; The target domain-related load knowledge and the target domain regional feature vector are fused together. The fusion ratio of source domain knowledge and target domain features is dynamically adjusted through a gating mechanism to obtain preliminary target domain load features. The initial target domain load characteristics are reconstructed using a region-specific recovery network. The characteristics are adjusted according to the industrial structure and electricity consumption habits of the target domain to obtain the target domain load characteristics.

[0011] Optionally, the step of fusing and fine-tuning the historical load data of the target domain with the load characteristics of the target domain to obtain optimized load characteristics of the target domain, and then performing time-series prediction processing on the optimized load characteristics of the target domain according to a multi-step prediction algorithm to generate cross-regional load prediction results includes: The historical load data of the target domain is time-aligned with the load characteristics of the target domain through a time alignment mechanism to obtain time-aligned load data feature pairs. Based on the time-aligned load data feature pairs, the local pattern information of the target domain historical load data and the target domain load features are fused through the residual connection network. The contribution weights of historical data and migration features are dynamically adjusted to obtain the fused target domain load features. The fused target domain load characteristics are processed for region-specific adjustment using a parameter fine-tuning network. The characteristics are then finely corrected according to the actual load variation pattern of the target domain to obtain optimized target domain load characteristics. The optimized target domain load characteristics are input into a multi-step time series prediction network for recursive prediction processing. The prediction result of each step is used as the input feature for the next step. The error propagation of multi-step prediction is suppressed by an error accumulation control mechanism to obtain cross-regional load prediction results.

[0012] Optionally, the fused target domain load characteristics are subjected to region-specific adjustment processing through a parameter fine-tuning network, and the characteristics are fine-grainedly corrected according to the actual load change pattern of the target domain to obtain optimized target domain load characteristics, including: Based on the fused target domain load characteristics, the load change patterns of the target domain during weekday morning peak, evening peak, midday trough, and late-night trough are extracted, and the average load and load change rate for each period are calculated to obtain the target domain period load pattern parameters. Based on the target domain time period load pattern parameters, the load characteristics of each time period are converted into weight adjustment coefficients of the fine-tuning network through a parameter mapping function. The adjustment coefficient for daytime production in industrial areas ranges from 1.2 to 1.5, and the adjustment coefficient for morning and evening peak electricity consumption in residential areas ranges from 1.3 to 1.6, thus obtaining the regionalized weight adjustment coefficients. The fused target domain load characteristics and the regionalized weight adjustment coefficient input parameter fine-tuning network are used for feature correction processing. Based on the actual industrial structure ratio of the target domain and residents' electricity consumption habits, the feature values ​​of different time periods are adjusted differently to obtain time-differentiated load characteristics. The time-varying load characteristics are corrected for temperature sensitivity using a seasonal correction function. Based on historical data of summer cooling load and winter heating load in the target domain, the characteristics of the corresponding seasons are adjusted to obtain the optimized target domain load characteristics.

[0013] Secondly, this application provides an intelligent power load forecasting system, the intelligent power load forecasting system comprising: The quantization module is used to perform regional difference quantization processing on the load data of the source domain and the target domain through the regional load pattern recognition algorithm, so as to obtain a regional feature vector containing load rate, peak-valley difference rate and load fluctuation coefficient; The calculation module is used to perform temporal encoding and attention weight calculation on the regional feature vector to obtain the source domain load temporal features and load attention weight distribution; The processing module is used to perform domain adaptation processing on the source domain load time-series characteristics through domain adversarial training to obtain domain-invariant load characteristics; The migration module is used to perform cross-domain migration processing on the domain-invariant load features and load attention weight distribution according to the knowledge fusion algorithm to obtain the target domain load features; The prediction module is used to fuse and fine-tune the historical load data of the target domain with the load characteristics of the target domain to obtain optimized load characteristics of the target domain. Then, it performs time-series prediction processing on the optimized load characteristics of the target domain according to a multi-step prediction algorithm to generate cross-regional load prediction results.

[0014] Thirdly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described intelligent power load prediction method.

[0015] The technical solution provided in this application quantifies the regional differences in load data of the source and target domains through a regional load pattern recognition algorithm. This transforms abstract regional differences into specific numerical feature vectors containing load factor, peak-to-valley difference rate, and load fluctuation coefficient, achieving standardized representation of load characteristics in different regions and laying a quantitative foundation for subsequent cross-domain knowledge transfer. Temporal coding and attention weight calculation fully exploit the temporal dependencies of load data. The self-attention mechanism automatically identifies the correlation strength between different time steps, solving the problem that traditional methods struggle to capture long-distance temporal dependencies, enabling the prediction model to more accurately grasp load change patterns. Domain adversarial training, through adversarial learning between the feature extraction network and the region discriminator, achieves the transformation of source domain feature distribution into domain-invariant features. This algorithm innovatively utilizes gradient inversion... The transfer technique eliminates region-specific biases, ensuring that the extracted load features retain the essential information needed for prediction without region-specific interference, significantly improving the generalization ability of the features. The knowledge fusion algorithm combines load attention weight distribution for intelligent knowledge screening, automatically identifying knowledge fragments related to the target domain based on attention weight values, avoiding the negative impact of irrelevant knowledge. The gating mechanism dynamically adjusts the fusion ratio, reflecting the algorithm's adaptability and intelligently adjusting the dependence on source domain knowledge according to the richness of the target domain data. The fusion fine-tuning process deeply integrates historical load data of the target domain with the transferred features, achieving an organic combination of global knowledge and local characteristics through residual connection networks and parameter fine-tuning networks. The multi-step prediction algorithm employs recursive prediction and error accumulation control mechanisms, effectively suppressing the error propagation problem in long-term prediction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of one embodiment of the intelligent power load prediction method in this application. Figure 2 This is a schematic diagram of one embodiment of the intelligent power load prediction system in this application. Detailed Implementation

[0018] This application provides a method, system, and storage medium for intelligent prediction of electricity load. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0019] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent power load prediction method in this application includes: Step S101: Perform regional difference quantification on the load data of the source domain and the target domain using the regional load pattern recognition algorithm to obtain a regional feature vector containing load rate, peak-valley difference rate and load fluctuation coefficient; Step S102: Perform temporal encoding and attention weight calculation on the regional feature vectors to obtain the temporal features of the source domain load and the load attention weight distribution; Step S103: Perform domain adaptation processing on the temporal characteristics of the source domain load through domain adversarial training to obtain domain-invariant load characteristics; Step S104: Perform cross-domain transfer processing on the domain-invariant load characteristics and load attention weight distribution according to the knowledge fusion algorithm to obtain the target domain load characteristics; Step S105: The historical load data of the target domain is fused and fine-tuned with the load characteristics of the target domain to obtain the optimized load characteristics of the target domain. The optimized load characteristics of the target domain are then subjected to time-series prediction processing according to the multi-step prediction algorithm to generate cross-regional load prediction results.

[0020] It is understood that the executing entity of this application can be an intelligent power load forecasting system, a terminal, or a server; no specific limitation is made here. This application's embodiment uses a server as an example for illustration.

[0021] Specifically, when performing regional difference quantification on load data in the source and target domains using a regional load pattern recognition algorithm, historical electricity load data is preprocessed to remove outliers caused by equipment failures or communication anomalies. Missing data points are then filled using linear interpolation, and after standardization, continuous and complete load time-series data is obtained. Based on the standardized load time-series data, three key indicators of the regional feature vector are calculated: the load factor, obtained as the ratio of the average load value to the maximum load value, reflects the load utilization level of the region; the peak-valley difference rate, calculated by dividing the difference between the peak load and the valley load by the average load value, reflects the fluctuation range of electricity load; and the load fluctuation coefficient, determined by the ratio of the load standard deviation to the average load value, characterizes the stability of the load.

[0022] When the regional feature vector is processed by time-series encoding, a sliding window mechanism is used to divide continuous load data into fixed-length sequence segments, each containing 24 hours of load data points. Multi-scale time-series decomposition is performed on each sequence segment to extract short-term fluctuation components to capture hourly load changes, daily cycle components to identify daily electricity consumption patterns, and long-term trend components to reflect seasonal variations. A location encoder generates a unique location identifier for each time step, which is superimposed with the multi-scale time-series representation to form the source domain load time-series feature. A self-attention mechanism calculates the correlation between different time steps, generating a time-series correlation matrix, which is then normalized to obtain the load attention weight distribution.

[0023] During domain adversarial training, both the temporal features of the source domain load and the load data of the target domain are simultaneously input into the load feature extraction network. The network extracts deep feature representations through multi-layer convolution and pooling operations. The region discriminator attempts to identify which region the features originate from, while the gradient inversion layer backpropagates the gradient signal of the discriminator, forcing the feature extraction network to generate region-independent feature representations. The training process continues until the distance between load patterns in different regions is less than a preset threshold. At this point, the obtained domain-invariant load features retain the essential laws of load variation while eliminating region-specific interference factors.

[0024] The knowledge fusion process compresses domain-invariant load characteristics into common load knowledge through a knowledge distillation network, and selects knowledge fragments relevant to the target domain based on the load attention weight distribution. A gating mechanism dynamically adjusts the fusion ratio of source domain knowledge and target domain features; when historical data for the target domain is abundant, the weight of target domain features is increased; when data is scarce, more reliance is placed on source domain knowledge. A region-specific recovery network adjusts the features according to the industrial structure and electricity consumption habits of the target domain, generating load characteristics that conform to the characteristics of the target domain.

[0025] The fusion fine-tuning process aligns historical load data of the target domain with its load characteristics over time. The residual connection network extracts local pattern information from the historical data and fuses it with migration features. The parameter fine-tuning network performs fine-grained corrections based on the actual load variation patterns of the target domain, particularly differentiating adjustments for load characteristics during typical weekday peak hours and midday off-peak periods. The multi-step time-series forecasting network employs a recursive forecasting approach, using the current forecast result as input for the next step. An error accumulation control mechanism prevents the gradual amplification of forecast errors, generating accurate cross-regional load forecast results.

[0026] In one specific embodiment, the process of performing step S101 may specifically include the following steps: Outlier detection and missing value interpolation are performed on the historical electricity load data of the source and target domains to obtain standardized load time-series data. The average load value and the maximum load value are calculated based on the standardized load time series data. The ratio of the average load value to the maximum load value is then calculated to obtain the load factor value. Peak load and valley load are extracted based on the standardized load time series data. The difference between peak load and valley load is divided by the average load value to calculate the peak-valley difference rate. The load standard deviation is calculated based on the standardized load time series data. The load standard deviation is then compared with the average load value to obtain the load fluctuation coefficient. The load rate, peak-valley difference rate, and load fluctuation coefficient are combined into a vector to obtain the regional feature vector.

[0027] Specifically, outlier detection in historical electricity load data employs a statistically based box plot method. The first and third quartiles of the load data are calculated to determine the interquartile range. Data points exceeding the range of the first quartile minus 1.5 times the interquartile range or the third quartile plus 1.5 times the interquartile range are marked as outliers. Outliers can be caused by data acquisition equipment malfunctions, communication transmission errors, or sudden power grid events. These outlier data points can severely impact the accuracy of subsequent load characteristic calculations. Missing value imputation uses a linear interpolation method, filling in missing values ​​based on the effective load values ​​before and after the missing point, linearly calculating the time interval to ensure the continuity and integrity of the load time series data. Standardization involves calculating the mean and standard deviation of all load data, subtracting the mean from each data point and dividing by the standard deviation, ensuring a uniform numerical range and distribution characteristics for load data across different time periods and regions. The average load value is calculated based on the standardized load time series data, by summing the load values ​​at all times within a specified time period and dividing by the total number of times. The maximum load value is directly extracted from the load sequence of that time period. Load factor, as an important indicator reflecting the utilization of power equipment, is calculated by dividing the average load value by the maximum load value. The closer the value is to 1, the more stable the load and the higher the equipment utilization rate.

[0028] Extracting peak and off-peak loads requires a comprehensive comparison of 24-hour load data. Peak load typically occurs during peak electricity consumption periods, while off-peak load occurs during the late night when electricity consumption is low. The peak-to-off-peak ratio is calculated by first determining the difference between the peak and off-peak loads, then dividing this difference by the average load value. The resulting value reflects the load fluctuation range throughout the day; a larger peak-to-off-peak ratio indicates more drastic intraday load changes. The load standard deviation is calculated using the statistical standard deviation formula: first, the difference between the load value at each moment and the average load value is calculated; these differences are squared, summed, divided by the total number of moments, and the square root is taken. The load fluctuation coefficient is obtained by dividing the load standard deviation by the average load value. This coefficient quantifies the degree of load dispersion around the average value; a smaller coefficient indicates a more stable load.

[0029] The combined processing of regional feature vectors arranges the calculated load factor, peak-valley difference rate, and load fluctuation coefficient in a fixed order to form a three-dimensional vector. These three indicators characterize the regional electricity load from different perspectives: the load factor reflects equipment utilization efficiency, the peak-valley difference rate reflects the intraday load variation pattern, and the load fluctuation coefficient characterizes load stability. These three indicators complement each other, together forming a complete feature representation of the regional load pattern, solving the problem that traditional methods relying on a single indicator cannot comprehensively reflect the regional load characteristics.

[0030] In one specific embodiment, the process of performing step S102 may specifically include the following steps: The regional feature vectors are segmented into sequences according to time order using a sliding window mechanism to obtain fixed-length load feature sequence segments. Multi-scale time series decomposition processing is performed on the load characteristic sequence segment to extract short-term fluctuation components, daily cycle components and long-term trend components respectively. The three components are then fused to obtain a multi-scale time series representation. Based on multi-scale temporal representation, a position encoder generates a position encoding vector according to the relative position of each time step. The position encoding vector is then superimposed with the multi-scale temporal representation to obtain the temporal features of the source domain load. The correlation scores between different time steps of the source domain load time series characteristics are calculated using a self-attention mechanism to obtain the time series correlation matrix; The temporal correlation matrix is ​​normalized so that the sum of the attention weights at each time step is a unit value, thus obtaining the load attention weight distribution.

[0031] Specifically, when the sliding window mechanism segments the regional feature vectors, the window length is set to 24 time points, corresponding to the load data of 24 hours a day. The window slides forward one time point at a time. In the specific processing, the first window contains the regional feature vectors of time points 1 to 24, the second window contains the regional feature vectors of time points 2 to 25, and so on, until the entire time series is traversed. The regional feature vectors in each window are arranged in chronological order, forming a fixed-length load feature sequence segment. This sequence segment retains the temporal variation information of load rate, peak-to-valley difference rate, and load fluctuation coefficient over a continuous 24 hours. The multi-scale time series decomposition processing adopts the empirical mode decomposition method to decompose each load feature sequence segment into components of different time scales. The short-term fluctuation component is obtained by extracting high-frequency oscillations with a period of 2-4 hours, reflecting the rapid change characteristics of the load, such as the sharp fluctuations during the morning and evening peak electricity consumption periods. The daily periodic component is extracted by using a bandpass filter to extract the signal components of the 24-hour period, capturing the load change pattern within a day, including the periodic pattern of high load during the day and low load at night. The long-term trend component was obtained using a moving average method with a window length of 7 days, reflecting the seasonal variation and long-term growth trend of the load. The three components were fused using a weighted summation method, with the weights determined based on the contribution of each component to the load forecast: the short-term fluctuation component had a weight of 0.3, the daily cycle component had a weight of 0.5, and the long-term trend component had a weight of 0.2. The fused component yielded a time-series representation containing information from multiple time scales.

[0032] The position encoder generates a unique encoding vector based on the relative position of each time step in the sequence, using a combination of sine and cosine functions. Even-numbered dimensions use sine functions, while odd-numbered dimensions use cosine functions, with the frequency of the functions decreasing with each dimension, ensuring the distinctiveness of the encoding vectors at different positions. The position encoding vectors are then superimposed on the multi-scale temporal representations element-wise. This superposition process preserves the numerical information of the original temporal features while injecting positional information, enabling the subsequent attention mechanism to recognize the relative positional relationships between different time steps.

[0033] The self-attention mechanism first maps the temporal features of the source domain load into query vectors, key vectors, and value vectors. A relevance score is then calculated by performing a dot product operation between the query vector and all key vectors. The relevance score reflects the strength of the dependency between different time steps; a higher score indicates a more similar load pattern between two time steps. Each element of the temporal relevance matrix represents the original relevance score between two corresponding time steps, and the diagonal elements of the matrix represent the relevance of a time step to itself. Normalization involves applying a softmax function to each row of the temporal relevance matrix, converting the original relevance scores into a probability distribution. The normalization calculation first exponentially calculates all scores in each row, then divides by the sum of the exponents for that row, ensuring that the sum of the attention weights assigned to all other time steps for each time step equals 1. The normalized load attention weight distribution directly reflects which historical time steps should be prioritized when predicting the load at a given time step; time steps with larger weights contribute more to the current prediction. This mechanism enables the model to automatically learn temporal dependency patterns in the load data, solving the problem of traditional methods struggling to capture long-distance temporal dependencies.

[0034] In one specific embodiment, the process of executing step S103 may specifically include the following steps: The source domain load time series features are input into the load feature extraction network for deep feature extraction processing, and the target domain regional load data are input into the same load feature extraction network for feature extraction processing to obtain the source domain load deep features and the target domain load deep features. Based on the deep features of the load in the source domain and the load in the target domain, a region discriminator is used to classify the load source regions and calculate the region classification loss value. At the same time, a gradient inversion layer is used to backpropagate the gradient of the region classification loss to obtain an inverted gradient signal used to confuse regional differences. The parameters of the load feature extraction network are updated based on the inverted gradient signal, so that the load features generated by the load feature extraction network cannot be identified by the region discriminator as originating from the region, thus obtaining region-independent load confusion features. The load confusion features that are not related to the region are processed by calculating the load pattern distance between regions through the cross-regional load distribution difference measure. Training stops when the load pattern distance between regions is less than the preset convergence threshold, and the domain-invariant load features are obtained.

[0035] Specifically, the load feature extraction network employs a deep convolutional neural network architecture, specifically designed for feature extraction from temporal load data. The network's first convolutional layer has 64 convolutional kernels, each 1×3 in size, representing coverage of three consecutive time points. When the temporal load features from the source domain are input into the network, each convolutional kernel slides across the time series with a stride of 1. During convolution, the kernel's weights are multiplied by the corresponding input data, summed, and a bias term is added. The result is then processed through the ReLU activation function to obtain the feature map. The 64 convolutional kernels in the first layer learn different local temporal patterns, such as load increase, decrease, or stationary patterns. The second convolutional layer uses 128 1×3 kernels to further convolution the feature map output from the first layer, extracting more complex temporal combination patterns. The third convolutional layer has 256 kernels to further abstract the feature representation. Each convolutional layer is followed by batch normalization and ReLU activation. Batch normalization standardizes the features to a distribution with a mean of 0 and a variance of 1 by calculating the mean and variance of the current batch of data. Then, learnable scaling and translation parameters are used to restore the expressive power of the features. After three convolutional layers, global average pooling is used to compress feature sequences of different lengths into a fixed 256-dimensional vector. This vector undergoes a non-linear transformation through two fully connected layers (512 and 256 neurons respectively), ultimately outputting a 256-dimensional deep feature representation. The target domain region load data is processed using the exact same network structure and parameters, ensuring that the features of both domains are represented in the same feature space. The region discriminator is designed as a lightweight fully connected network. The first fully connected layer maps the 256-dimensional input features to a 128-dimensional hidden representation, using the Leaky ReLU activation function (with a slope of 0.2 on the negative part) to avoid neuron death. The second fully connected layer maps the 128-dimensional features to a 2-dimensional output, representing the original scores of the input features belonging to the source and target domains, respectively. The original scores are converted into a probability distribution using the softmax function, calculated by dividing the index value of each category by the sum of the index values ​​of all categories. During discriminator training, the true labels corresponding to the source domain features are one-hot vectors of [1,0], and the target domain features correspond to [0,1]. The region classification loss uses the cross-entropy loss function, calculating the difference between the predicted probability distribution and the true label distribution; the loss value is equal to the negative logarithm of the predicted probability corresponding to the true category. The gradient reversal layer keeps the input unchanged during forward propagation, but multiplies the gradient by -λ (λ is the gradient reversal coefficient, initially set to 0.1, gradually increasing to 1.0 as training progresses) during backpropagation, achieving a complete reversal of the gradient direction.

[0036] The parameter update of the feature extraction network involves balancing two optimization objectives. The gradient generated by the load forecasting task guides the network to learn feature representations useful for load forecasting; this gradient is calculated using the standard backpropagation algorithm, propagating layer by layer from the prediction error. The gradient generated by the region discrimination task, after inversion, guides the network to generate features that confuse the discriminator. Specifically, the update process first calculates the gradient of the load forecasting loss with respect to the network parameters, then calculates the gradient of the region classification loss with respect to the network parameters and multiplies it by -λ. The two gradients are then added together with preset weights to obtain the final gradient. The Adam optimizer is used to update the network parameters based on the final gradient, with a learning rate of 0.001 and momentum parameters β1=0.9 and β2=0.999. After each parameter update, the features generated by the feature extraction network gradually become both informative of the load forecasting requirements and difficult for the discriminator to identify the source region.

[0037] Cross-regional load distribution difference measurement uses a kernel method to calculate the distance between the feature distributions of two domains. First, a Gaussian radial basis function is chosen as the kernel function, and the bandwidth parameter is determined heuristically using the median, equal to the median of the Euclidean distances between all feature pairs. For n feature samples from the source domain and m feature samples from the target domain, the sum of the kernel function values ​​for each intra-domain sample pair is divided by the square of the number of samples to obtain the average kernel value for the source and target domains. Then, the sum of the kernel function values ​​for cross-domain sample pairs is divided by the product of the number of samples to obtain the cross-domain average kernel value. The maximum mean difference is equal to the sum of the intra-domain average kernel values ​​minus twice the cross-domain average kernel value, then taking the square root. During training, the distribution difference is calculated after each epoch. When the difference value is less than 0.01 for five consecutive epochs, the convergence condition is considered met. The feature representation at this point is the domain-invariant load feature, successfully eliminating the distribution difference between the high base load of the source domain industrial area and the large load fluctuation of the target domain commercial area.

[0038] In one specific embodiment, the process of executing step S104 may specifically include the following steps: Domain-invariant load features are compressed using a knowledge distillation network to obtain compressed common load knowledge. Based on the load attention weight distribution, common load knowledge is weighted and selected. Knowledge fragments with high relevance to the target domain load pattern are selected according to the attention weight value to obtain target domain relevant load knowledge. The target domain load knowledge is fused with the target domain regional feature vector. The fusion ratio of source domain knowledge and target domain features is dynamically adjusted through a gating mechanism to obtain preliminary target domain load characteristics. The initial target domain load characteristics are reconstructed using a region-specific recovery network. The characteristics are adjusted according to the industrial structure and electricity consumption habits of the target domain to obtain the target domain load characteristics.

[0039] Specifically, the knowledge distillation network compresses domain-invariant load features using a teacher-student architecture. The teacher network retains the original 256-dimensional feature representation, while the student network is designed with a more compact structure, containing two fully connected layers. The first layer maps the 256-dimensional input to 128 dimensions, and the second layer further compresses it to 64 dimensions. During distillation, the teacher network's output is processed by a softmax function with a temperature parameter of 3 to generate soft labels, and the student network learns to match the distribution of these soft labels. The temperature parameter smooths the probability distribution, allowing the student network to learn more subtle knowledge structures from the teacher network. During training, two loss functions are optimized simultaneously: the KL divergence loss between the student network output and the teacher network's soft labels, and the prediction loss of the student network on the original task, with a weight ratio of 0.7:0.3. After distillation, the 64-dimensional compressed features retain the most relevant common knowledge for load prediction from the original 256-dimensional features while removing redundant information.

[0040] The weighted selection of common load knowledge by the load attention weight distribution is achieved through element-wise multiplication, multiplying each dimension of the 64-dimensional common load knowledge vector by the attention weight of the corresponding time step. The attention weight values ​​reflect the importance of different time steps to the current prediction; time steps with higher weights indicate a stronger correlation between their load patterns and the current time in the target domain. The selection process sets a threshold of 1.5 times the average attention weight, retaining only knowledge fragments with weights exceeding the threshold and setting dimensions with weights below the threshold to zero. This selection mechanism ensures that the knowledge transferred to the target domain is highly correlated with the target domain's load patterns, avoiding interference from irrelevant knowledge. The selected knowledge fragments are reorganized into continuous feature vectors, forming target domain-relevant load knowledge.

[0041] The knowledge fusion processing employs a gated fusion unit to dynamically combine source domain knowledge and target domain features. The gating mechanism includes two gates: an update gate and a reset gate. The update gate is calculated using the sigmoid function, taking the concatenation of target domain relevant load knowledge and target domain regional feature vectors as input. Its output value, between 0 and 1, represents the proportion of target domain features retained. The reset gate also uses the sigmoid function to control the degree of utilization of source domain knowledge. During fusion calculation, the reset gate output is first multiplied by the target domain relevant load knowledge to obtain filtered source domain information. This is then added to the target domain regional feature vector and activated using the tanh activation function to generate candidate features. The final fused features are obtained by weighted combination of the original target domain features and candidate features through the update gate. When historical data for the target domain is abundant, the update gate output is close to 1, primarily retaining the target domain's own features; when target domain data is scarce, the update gate output is smaller, relying more on source domain knowledge. The region-specific recovery network consists of three sub-networks: an industry structure encoder, an electricity consumption habit encoder, and a feature reconstructor. The industry structure encoder encodes the proportion of industrial, commercial, and residential electricity consumption in the target domain into a 32-dimensional vector. One-hot encoding is used to represent the dominant industry type, which is then mapped to the feature space through a fully connected layer. The electricity consumption habit encoder processes typical daily load curves of the target domain, extracting load features for key periods such as morning peak hours, evening peak hours, and midday off-peak hours, and encoding them into a 32-dimensional vector. The feature reconstructor concatenates the initial target domain load features with the two encoded vectors, performing a nonlinear transformation through a three-layer fully connected network. Residual connections between each layer maintain information flow. During reconstruction, the network learns to adjust the load baseline value based on the industry structure of the target domain and adjust the feature weights for peak and off-peak periods based on electricity consumption habits. The final output target domain load features not only include general load forecasting knowledge transferred from the source domain but also restore the unique load pattern features of the target domain, solving the problem of regional feature loss caused by direct transfer.

[0042] In one specific embodiment, the process of executing step S105 may specifically include the following steps: The historical load data of the target domain is time-aligned with the load characteristics of the target domain through a time alignment mechanism to obtain time-aligned load data feature pairs. Based on time-aligned load data feature pairs, the local pattern information of historical load data in the target domain is fused with the target domain load features through a residual connection network. The contribution weights of historical data and migration features are dynamically adjusted to obtain the fused target domain load features. The fused target domain load characteristics are processed for region-specific adjustment through a parameter fine-tuning network. The characteristics are then finely corrected according to the actual load change pattern of the target domain to obtain the optimized target domain load characteristics. The optimized target domain load characteristics are input into a multi-step time series prediction network for recursive prediction processing. The prediction result of each step is used as the input feature for the next step. The error propagation of multi-step prediction is suppressed by an error accumulation control mechanism to obtain cross-regional load prediction results.

[0043] Specifically, the time alignment mechanism uses a timestamp matching algorithm to precisely map historical load data of the target domain to load characteristics of the target domain. Historical load data of the target domain is stored at an hourly granularity, with each data point containing a timestamp and a corresponding load value. After the aforementioned processing, the load characteristics of the target domain form a time-series feature sequence, with each feature vector corresponding to a specific time window. The alignment process first extracts the timestamps from the historical load data and converts them to a unified time format. Then, based on the starting point of the time window in the feature sequence, it searches for the load record for the corresponding time period in the historical data. When the sampling frequency of the historical data is inconsistent with the feature sequence, a linear interpolation method is used to generate the load value for the missing time point. After alignment, each time point contains the actual historical load value and the corresponding feature vector, forming a time-aligned load data feature pair.

[0044] When processing the fusion process using the residual connection network, local pattern information is first extracted from time-aligned load data feature pairs. Local patterns are obtained using a sliding window approach, with a window size of 6 hours and a sliding window of 1 hour. The mean, variance, and trend of the load data within the window are calculated. These local statistics, along with the target domain load features corresponding to the time window, are input into the residual block. The residual block contains two paths: the main path performs a nonlinear transformation on the input through two fully connected layers, while the shortcut path directly passes the original input. The first layer of the main path maps the input to a higher-dimensional space for feature interaction, and the second layer compresses it back to the original dimension. Dynamic weights are calculated using an attention mechanism, automatically adjusted based on the completeness and reliability of historical data; higher weights are assigned to high-quality historical data, while their contribution is reduced when there is significant data loss. The outputs of the two paths are weighted and summed to obtain the residual block output. Multiple residual blocks are then concatenated to output the fused target domain load features.

[0045] The parameter fine-tuning network performs feature correction based on the actual load variation patterns of the target domain. First, it analyzes recent load data of the target domain to extract typical daily load curves. A clustering algorithm categorizes historical load curves into three types: weekdays, weekends, and holidays, calculating representative load patterns for each type. The fine-tuning network contains three parallel correction branches, each handling feature adjustments for different types of days. Each branch learns feature transformation parameters for that type of day through a fully connected layer, including peak period offset, peak-to-valley ratio adjustment coefficient, and base load correction value. The fine-grained correction process inputs the fused target domain load features into the corresponding correction branch, selecting appropriate transformation parameters based on the type of the current predicted day, applying different degrees of adjustment to different dimensions of the features. The corrected features are then stabilized within their numerical range using a batch normalization layer, outputting the optimized target domain load features.

[0046] The multi-step time-series prediction network adopts a sequence-to-sequence architecture. The encoder processes the optimized target domain load features, and the decoder generates load predictions for multiple future time steps. The recursive prediction process starts from the first prediction time step. The decoder generates the prediction value for the current time step based on the encoder output and previous historical data. This prediction value, along with its corresponding uncertainty estimate, serves as the input feature for the next time step, continuing to generate subsequent predictions. The error accumulation control mechanism is implemented through three strategies: First, a confidence interval is calculated for each prediction step, and the weight of historical data is increased when the accumulated uncertainty exceeds a threshold; second, a teacher-forced technique is used, randomly using the true value instead of the predicted value as the input for the next step during training to reduce error propagation; third, a decay factor is introduced, assigning a smaller loss weight to the prediction error of more distant time steps. The prediction process continues until the preset prediction duration is reached, ultimately outputting a sequence of cross-regional load prediction results for the next 24 hours or longer.

[0047] In one specific embodiment, the process of performing region-specific adjustment of the fused target domain load characteristics using a parameter fine-tuning network can specifically include the following steps: Based on the fused target domain load characteristics, the load change patterns of the target domain during weekday morning peak, evening peak, midday off-peak, and late-night off-peak periods are extracted, and the average load and load change rate for each period are calculated to obtain the target domain period load pattern parameters. Based on the target domain time period load pattern parameters, the load characteristics of each time period are converted into weight adjustment coefficients of the fine-tuning network through the parameter mapping function. The adjustment coefficient for daytime production in industrial areas ranges from 1.2 to 1.5, and the adjustment coefficient for morning and evening peak electricity consumption in residential areas ranges from 1.3 to 1.6, thus obtaining the regionalized weight adjustment coefficients. The merged target domain load characteristics and regionalized weight adjustment coefficients are used to fine-tune the network for feature correction. Based on the actual industrial structure ratio of the target domain and residents' electricity consumption habits, the feature values ​​for different time periods are adjusted differently to obtain time-differentiated load characteristics. The time-varying load characteristics are corrected for temperature sensitivity using a seasonal correction function. Based on historical data of summer cooling load and winter heating load in the target domain, the characteristics of the corresponding seasons are adjusted to obtain the optimized load characteristics of the target domain.

[0048] Specifically, the time-period load pattern parameter extraction process segments the fused target domain load characteristics according to predefined time periods. Weekday morning peak is defined as 7:00 AM to 9:00 AM, evening peak as 5:00 PM to 8:00 PM, midday trough as 12:00 PM to 2:00 PM, and late-night trough as 2:00 AM to 4:00 AM. For the load characteristic data within each time period, the arithmetic mean of the load values ​​at all times within that time period is first calculated as the time-period load average. The load change rate is obtained by calculating the difference between load values ​​at adjacent times within the time period, dividing by the time interval, and then averaging all instantaneous change rates to obtain the time-period average load change rate. The load change rate during the morning peak period is usually positive, reflecting a rapid increase in load; the load change rate during the late-night trough period is close to zero, indicating that the load remains stable at a low level. These parameters combine to form the target domain time-period load pattern parameter matrix, with each row representing a time period and containing the load average and change rate information for that period.

[0049] The parameter mapping function converts the time-period load pattern parameters into weight adjustment coefficients for the fine-tuning network based on the target domain's regional type and time-period characteristics. The mapping function employs a piecewise linear transformation. For industrial areas, when the average load during daytime production hours (8:00 AM to 5:00 PM) is between 70% and 85% of the area's historical peak load, the adjustment coefficient linearly increases from 1.2 to 1.5. When the average load is below 70%, the adjustment coefficient remains at 1.2; when it is above 85%, the adjustment coefficient remains at 1.5. The mapping rules for residential areas consider the characteristics of morning and evening peak electricity consumption. During the morning peak (6:00 AM to 8:00 AM) and evening peak (6:00 PM to 10:00 PM), when the average load reaches 75% to 90% of the area's historical peak load, the adjustment coefficient increases from 1.3 to 1.6. The mapping process also considers the load change rate; periods with a higher change rate are assigned a higher adjustment coefficient to ensure that rapidly changing periods receive sufficient feature enhancement.

[0050] The parameter fine-tuning network receives the fused target domain load characteristics and regionalized weight adjustment coefficients, and performs feature correction through matrix operations. The network first divides the input features by time period, multiplying the feature vector of each time period by the corresponding adjustment coefficient. For industrial areas, the feature values ​​during daytime production periods are multiplied by a coefficient between 1.2 and 1.5 to emphasize production load characteristics; during nighttime non-production periods, the original value is maintained or slightly reduced. For residential areas, the focus is on morning and evening peak hours, with corresponding features multiplied by a coefficient between 1.3 and 1.6 to highlight residential electricity consumption patterns. The industrial structure ratio is reflected through weighted coefficients; areas with a high industrial proportion have a larger weight for industrial time period adjustment coefficients, while areas with a high residential proportion have a dominant adjustment coefficient for residential time periods. The differentiated adjustment process is completed through element-wise multiplication and weighted summation, outputting time-differentiated load characteristics.

[0051] The seasonal correction function handles the characteristic adjustment of temperature-sensitive loads. First, it extracts typical patterns of summer cooling load and winter heating load from historical data. Summer cooling load is obtained by analyzing the correlation between daily maximum temperature and daily maximum load from June to September. When the temperature exceeds 26 degrees Celsius, the load increase for each 1-degree increase in temperature is calculated using linear regression. Winter heating load is analyzed from November to February of the following year, examining the load increment pattern when the temperature is below 10 degrees Celsius. The seasonal correction process calculates the degree of temperature deviation from the comfort range based on the current season and forecast temperature, converting this deviation into a characteristic adjustment amount through a correction function. The correction function uses an S-shaped curve, saturating the adjustment amount at extreme temperatures to avoid over-correction. The corresponding dimensions of the time-varying load characteristics are added to or subtracted from the temperature correction amount to complete the seasonal correction, outputting the final optimized target domain load characteristics.

[0052] The above describes the intelligent power load prediction method in the embodiments of this application. The following describes the intelligent power load prediction system in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the intelligent power load prediction system in this application includes: The quantization module is used to perform regional difference quantization processing on the load data of the source domain and the target domain through the regional load pattern recognition algorithm, so as to obtain a regional feature vector containing load rate, peak-valley difference rate and load fluctuation coefficient; The calculation module is used to perform temporal encoding and attention weight calculation on the regional feature vector to obtain the source domain load temporal features and load attention weight distribution; The processing module is used to perform domain adaptation processing on the source domain load time-series characteristics through domain adversarial training to obtain domain-invariant load characteristics; The migration module is used to perform cross-domain migration processing on the domain-invariant load features and load attention weight distribution according to the knowledge fusion algorithm to obtain the target domain load features; The prediction module is used to fuse and fine-tune the historical load data of the target domain with the load characteristics of the target domain to obtain optimized load characteristics of the target domain. Then, it performs time-series prediction processing on the optimized load characteristics of the target domain according to a multi-step prediction algorithm to generate cross-regional load prediction results.

[0053] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the intelligent power load prediction method.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0055] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part 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 smart load forecasting 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 the present 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.

[0056] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent prediction of electricity load, characterized in that, The method includes: The regional load pattern recognition algorithm is used to quantify the regional differences in load data of the source and target domains to obtain a regional feature vector containing load factor, peak-valley difference rate and load fluctuation coefficient. The region feature vectors are processed by temporal encoding and attention weight calculation to obtain the source domain load temporal features and load attention weight distribution. Domain-adaptive processing is performed on the temporal characteristics of the source domain load through domain adversarial training to obtain domain-invariant load characteristics. The target domain load characteristics are obtained by performing cross-domain transfer processing on the domain-invariant load characteristics and load attention weight distribution based on the knowledge fusion algorithm. The target domain's historical load data is fused and fine-tuned with the target domain's load characteristics to obtain optimized target domain load characteristics. The optimized target domain load characteristics are then subjected to time-series prediction processing using a multi-step prediction algorithm to generate cross-regional load prediction results.

2. The intelligent load forecasting method according to claim 1, characterized in that, The method involves quantifying the regional differences in load data between the source and target domains using a regional load pattern recognition algorithm to obtain a regional feature vector containing load factor, peak-to-valley difference rate, and load fluctuation coefficient. Outlier detection and missing value interpolation are performed on the historical electricity load data of the source and target domains to obtain standardized load time-series data. Based on the standardized load time-series data, the average load value and the maximum load value are calculated, and the ratio of the average load value to the maximum load value is calculated to obtain the load rate value. Based on the standardized load time series data, peak load and valley load are extracted. The difference between peak load and valley load is divided by the average load value to calculate the peak-valley difference rate, and the peak-valley difference rate value is obtained. The load standard deviation is calculated based on the standardized load time series data. The load standard deviation is then compared with the average load value to obtain the load fluctuation coefficient. The load rate, peak-valley difference rate, and load fluctuation coefficient are then combined into a vector to obtain the regional feature vector.

3. The intelligent load forecasting method according to claim 1, characterized in that, The step of performing temporal encoding and attention weight calculation on the regional feature vector to obtain the source domain load temporal features and load attention weight distribution includes: The region feature vectors are segmented in chronological order using a sliding window mechanism to obtain load feature sequence segments of fixed length. The load characteristic sequence segment is subjected to multi-scale time series decomposition processing to extract short-term fluctuation components, daily cycle components and long-term trend components respectively. The three components are then fused to obtain a multi-scale time series representation. Based on the multi-scale temporal representation, a position encoder generates a position encoding vector according to the relative position of each time step. The position encoding vector is then superimposed with the multi-scale temporal representation to obtain the source domain load temporal features. The correlation scores between different time steps of the source domain load time series features are calculated using a self-attention mechanism to obtain the time series correlation matrix; The temporal correlation matrix is ​​normalized so that the sum of the attention weights at each time step is a unit value, thus obtaining the load attention weight distribution.

4. The intelligent load prediction method according to claim 1, characterized in that, The process of performing domain adaptation processing on the source domain load time-series features through domain adversarial training to obtain domain-invariant load features includes: The source domain load time-series features are input into the load feature extraction network for deep feature extraction processing, and the target domain regional load data are input into the same load feature extraction network for feature extraction processing to obtain the source domain load deep features and the target domain load deep features. Based on the deep features of the source domain load and the deep features of the target domain load, the load source region is classified by the region discriminator, the region classification loss value is calculated, and the gradient inversion layer is used to backpropagate the gradient of the region classification loss to obtain the inverted gradient signal used to confuse regional differences. The parameters of the load feature extraction network are updated according to the inverted gradient signal, so that the load features generated by the load feature extraction network cannot be identified by the region discriminator as their source region, thus obtaining region-independent load confusion features. The region-independent load confusion features are processed by calculating the inter-regional load pattern distance through cross-regional load distribution difference measurement. Training stops when the inter-regional load pattern distance is less than a preset convergence threshold, thus obtaining the domain-invariant load features.

5. The intelligent load prediction method according to claim 1, characterized in that, The process of performing cross-domain transfer processing on the domain-invariant load features and load attention weight distribution based on a knowledge fusion algorithm to obtain target domain load features includes: The domain-invariant load features are compressed using a knowledge distillation network to obtain compressed common load knowledge. Based on the load attention weight distribution, the common load knowledge is weighted and selected, and knowledge fragments with high relevance to the target domain load pattern are selected according to the attention weight value to obtain target domain related load knowledge; The target domain-related load knowledge and the target domain regional feature vector are fused together. The fusion ratio of source domain knowledge and target domain features is dynamically adjusted through a gating mechanism to obtain preliminary target domain load features. The initial target domain load characteristics are reconstructed using a region-specific recovery network. The characteristics are adjusted according to the industrial structure and electricity consumption habits of the target domain to obtain the target domain load characteristics.

6. The intelligent load forecasting method according to claim 1, characterized in that, The process of fusing and fine-tuning the historical load data of the target domain with the load characteristics of the target domain to obtain optimized load characteristics of the target domain, and then performing time-series prediction processing on the optimized load characteristics of the target domain according to a multi-step prediction algorithm to generate cross-regional load prediction results includes: The historical load data of the target domain is time-aligned with the load characteristics of the target domain through a time alignment mechanism to obtain time-aligned load data feature pairs. Based on the time-aligned load data feature pairs, the local pattern information of the target domain historical load data and the target domain load features are fused through the residual connection network. The contribution weights of historical data and migration features are dynamically adjusted to obtain the fused target domain load features. The fused target domain load characteristics are processed for region-specific adjustment using a parameter fine-tuning network. The characteristics are then finely corrected according to the actual load variation pattern of the target domain to obtain optimized target domain load characteristics. The optimized target domain load characteristics are input into a multi-step time series prediction network for recursive prediction processing. The prediction result of each step is used as the input feature for the next step. The error propagation of multi-step prediction is suppressed by an error accumulation control mechanism to obtain cross-regional load prediction results.

7. The intelligent load prediction method according to claim 6, characterized in that, The fused target domain load features are then subjected to region-specific adjustment processing via a parameter fine-tuning network. Based on the actual load variation patterns of the target domain, the features are fine-grained corrected to obtain optimized target domain load features, including: Based on the fused target domain load characteristics, the load change patterns of the target domain during weekday morning peak, evening peak, midday trough, and late-night trough are extracted, and the average load and load change rate for each period are calculated to obtain the target domain period load pattern parameters. Based on the target domain time period load pattern parameters, the load characteristics of each time period are converted into weight adjustment coefficients of the fine-tuning network through a parameter mapping function. The adjustment coefficient for daytime production in industrial areas ranges from 1.2 to 1.5, and the adjustment coefficient for morning and evening peak electricity consumption in residential areas ranges from 1.3 to 1.6, thus obtaining the regionalized weight adjustment coefficients. The fused target domain load characteristics and the regionalized weight adjustment coefficient input parameter fine-tuning network are used for feature correction processing. Based on the actual industrial structure ratio of the target domain and residents' electricity consumption habits, the feature values ​​of different time periods are adjusted differently to obtain time-differentiated load characteristics. The time-varying load characteristics are corrected for temperature sensitivity using a seasonal correction function. Based on historical data of summer cooling load and winter heating load in the target domain, the characteristics of the corresponding seasons are adjusted to obtain the optimized target domain load characteristics.

8. An intelligent power load forecasting system, characterized in that, For implementing the intelligent load forecasting method as described in any one of claims 1-7, the intelligent load forecasting system comprises: The quantization module is used to perform regional difference quantization processing on the load data of the source domain and the target domain through the regional load pattern recognition algorithm, so as to obtain a regional feature vector containing load rate, peak-valley difference rate and load fluctuation coefficient; The calculation module is used to perform temporal encoding and attention weight calculation on the regional feature vector to obtain the source domain load temporal features and load attention weight distribution; The processing module is used to perform domain adaptation processing on the source domain load time-series characteristics through domain adversarial training to obtain domain-invariant load characteristics; The migration module is used to perform cross-domain migration processing on the domain-invariant load features and load attention weight distribution according to the knowledge fusion algorithm to obtain the target domain load features; The prediction module is used to fuse and fine-tune the historical load data of the target domain with the load characteristics of the target domain to obtain optimized load characteristics of the target domain. Then, it performs time-series prediction processing on the optimized load characteristics of the target domain according to a multi-step prediction algorithm to generate cross-regional load prediction results.

9. A smart load forecasting device, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the intelligent load prediction method according to 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 run by the processor, it causes the processor to execute the intelligent load prediction method as described in any one of claims 1 to 7.

Images