Power demand prediction system and method based on deep fusion network

Abstract

The invention discloses a power demand prediction system and method based on a deep fusion network, and the method comprises the steps: carrying out the standardization preprocessing of multi-source data, introducing a self-adaptive weighting mechanism based on relation complexity, and dynamically fusing a plurality of statistical correlation coefficients to precisely recognize significant influence factors, the technical problem that traditional fixed weight screening is difficult to consider linear and nonlinear relationships is solved. Then, an Almong polynomial distribution lag model and coupling coordination degree analysis are utilized to quantify time lag contribution of influence factors and process feature interaction, and a final feature set including a dynamic conduction mechanism is constructed. And finally, through a deep fusion model integrating the convolutional neural network and the bidirectional long-short-term memory network, multi-scale local features and bidirectional long time sequence dependence are extracted in parallel, so that high-precision prediction of the power demand is realized, and the defects that an existing model is weak in generalization ability and difficult to capture a complex time sequence rule are overcome.

Description

Technical Field

[0001] This application relates to the field of power system planning and load forecasting technology, and more specifically, to a power demand forecasting system and method based on a deep fusion network. Background Technology

[0002] With the advancement of the dual-carbon strategy and the deepening of energy digital transformation, the power system's requirements for the accuracy of electricity demand forecasting are increasing. Accurate electricity demand forecasting is not only the cornerstone of grid security dispatch and power generation planning, but also an important basis for electricity market trading decisions and renewable energy consumption. Electricity load data itself has nonlinear, time-varying, and stochastic characteristics, and is deeply influenced by multi-dimensional external factors such as meteorological conditions, macroeconomic indicators, and social activities.

[0003] While existing technologies have evolved from traditional statistical methods to prediction models based on machine learning and deep learning, significant limitations remain in feature engineering and model architecture design. Specifically, in the screening of key influencing factors, current solutions typically introduce multiple statistical indicators such as Pearson, Kendall, Spearman, and the maximum information coefficient to measure the correlation between each factor and electricity load. However, when integrating multiple correlation indicators, these solutions generally employ an indiscriminate arithmetic averaging strategy, assuming that the weights of various indicators are fixed and equal. This approach ignores the essential differences in how different indicators capture variable relationships. For example, the Pearson coefficient strictly measures linear relationships, while the maximum information coefficient excels at discovering complex nonlinear relationships. In real-world scenarios, the relationship patterns between influencing factors and load are diverse. For instance, temperature and electricity consumption often exhibit a U-shaped nonlinear characteristic, while some economic indicators may be closer to a linear relationship. If a fixed-weight averaging method is used, in strongly nonlinear scenarios, lower linear coefficients will dilute higher nonlinear coefficients, causing the comprehensive indicator to fail to accurately reflect the true strong correlation. This lack of adaptive perception of relationship complexity directly leads to distortion in feature selection. Furthermore, in terms of model construction, existing single network structures are difficult to simultaneously take into account multi-scale local feature extraction and the capture of long-term and short-term bidirectional temporal dependencies, and often ignore the time lag effect of external factors on load and the coupling and coordination relationship between factors, thus limiting the generalization ability and prediction accuracy of the prediction system under complex working conditions.

[0004] Therefore, an optimized power demand forecasting scheme based on deep fusion networks is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a power demand forecasting system and method based on a deep fusion network.

[0006] According to one aspect of this application, a power demand forecasting method based on deep fusion networks is provided, comprising a training phase and a forecasting phase:

[0007] The training phase includes the following steps:

[0008] Obtain raw electricity consumption data and raw influencing factor data;

[0009] Data cleaning, downsampling, normalization, and dataset partitioning were performed on the original electricity consumption data and original influencing factor data to obtain normalized training and validation sets.

[0010] Key influencing factors were screened from the normalized training set to obtain a set of significant influencing factors.

[0011] Based on the electricity consumption data in the normalized training set, time lag correction and coupling relationship analysis are performed on the set of significant influencing factors to obtain the final feature set;

[0012] The deep fusion network is trained based on the final feature set and the normalized validation set to obtain a trained deep fusion network;

[0013] The prediction phase includes the following steps:

[0014] Obtain electricity consumption data and influencing factor data;

[0015] Extract the feature set to be predicted from electricity consumption data and influencing factor data;

[0016] The feature set to be predicted is input into a trained deep fusion network to obtain the predicted electricity demand value.

[0017] According to another aspect of this application, a power demand forecasting system based on a deep fusion network is provided, comprising: a training system and a forecasting system.

[0018] The training system includes:

[0019] The raw data acquisition module is used to acquire raw electricity consumption data and raw influencing factor data;

[0020] The module for obtaining normalized training sets and normalized validation sets is used to perform data cleaning, data downsampling, data normalization, and dataset partitioning on the original electricity consumption data and original influencing factor data to obtain normalized training sets and normalized validation sets.

[0021] The key influencing factor screening module is used to screen key influencing factors from the normalized training set to obtain a set of significant influencing factors.

[0022] The time delay correction and coupling relationship analysis module is used to perform time delay correction and coupling relationship analysis on the set of significant influencing factors based on the electricity consumption data in the normalized training set to obtain the final feature set;

[0023] The deep fusion network training module is used to train the deep fusion network based on the final feature set and the normalized validation set to obtain a trained deep fusion network.

[0024] The prediction system includes:

[0025] The data acquisition module is used to acquire electricity consumption data and influencing factor data;

[0026] The data extraction module is used to extract the feature set to be predicted from electricity consumption data and influencing factor data;

[0027] The electricity demand forecast acquisition module is used to input the feature set to be predicted into a trained deep fusion network to obtain the electricity demand forecast.

[0028] Compared with existing technologies, this method first performs standardized preprocessing on multi-source heterogeneous power and influencing factor data. Addressing the technical problem in multi-dimensional feature selection where fixed weights cannot simultaneously account for linear and nonlinear relationships, an adaptive weighting mechanism based on relational complexity is introduced to dynamically fuse multiple statistical correlation coefficients to accurately identify significant influencing factors. Subsequently, using the Almon multinomial distributed lag model and coupling coordination analysis, the time-lag contribution of influencing factors is quantified and the coupling interactions between features are handled, constructing a final feature set containing a dynamic transmission mechanism. Finally, by constructing a deep fusion model integrating convolutional neural networks and bidirectional long short-term memory networks, multi-scale local features and bidirectional long-term time-series dependencies are extracted in parallel. High-precision prediction of power demand is achieved through deep fusion with residual connections, overcoming the shortcomings of traditional models such as weak generalization ability and difficulty in capturing complex time-series patterns. Attached Figure Description

[0029] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0030] Figure 1 This is a flowchart of a power demand forecasting method based on a deep fusion network according to an embodiment of this application;

[0031] Figure 2 This is a data flow diagram illustrating a power demand forecasting method based on a deep fusion network according to an embodiment of this application.

[0032] Figure 3 This is a flowchart illustrating the process of data cleaning, data downsampling, data normalization, and dataset partitioning to obtain a normalized training set and a normalized validation set for a power demand forecasting method based on a deep fusion network, according to an embodiment of this application.

[0033] Figure 4 This is a flowchart illustrating the process of screening key influencing factors in a normalized training set to obtain a set of significant influencing factors, according to an embodiment of the present application, of a power demand forecasting method based on a deep fusion network.

[0034] Figure 5 This is a flowchart illustrating a method for predicting electricity demand based on a deep fusion network, according to an embodiment of this application, which involves fusing comprehensive correlation coefficients from various correlation coefficient vectors in a correlation coefficient vector set using adaptive weighting based on relational complexity to obtain a comprehensive correlation coefficient set.

[0035] Figure 6 This is a block diagram of a power demand forecasting system based on a deep fusion network according to an embodiment of this application. Detailed Implementation

[0036] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0037] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0038] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0039] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0040] Existing power demand forecasting technologies typically employ a fixed-weight strategy to fuse multiple correlation indicators when screening key influencing factors. This approach ignores the complexity differences between linear and nonlinear relationships between data, making it difficult to accurately identify significant features when facing complex and variable load characteristics, thus limiting the generalization ability and accuracy of the forecasting model. To address this issue and improve the robustness of feature selection and the accuracy of subsequent modeling, this application proposes a power demand forecasting method based on deep fusion networks. This scheme constructs an end-to-end forecasting framework from adaptive data screening to deep fusion of spatiotemporal features. Specifically, it first standardizes multi-source heterogeneous data and innovatively introduces an adaptive weighting mechanism based on relational complexity. This mechanism dynamically adjusts the fusion weights of linear and nonlinear correlation coefficients by quantifying the degree of nonlinearity between data, thereby intelligently identifying the true influencing factors. Furthermore, by combining the Almon multinomial distributed lag model and coupling coordination analysis, the selected features are corrected for time lags and their interaction relationships are reconstructed, effectively solving the problems of temporal misalignment and information redundancy among factors. Finally, by using a deep fusion model consisting of a one-dimensional convolutional neural network and a bidirectional long short-term memory network, multi-scale local features and bidirectional temporal dependencies are extracted in parallel, achieving high-precision and robust prediction of electricity demand.

[0041] Figure 1 This is a flowchart of a power demand forecasting method based on a deep fusion network according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating a power demand forecasting method based on a deep fusion network according to an embodiment of this application. Figure 1 and Figure 2As shown, an embodiment of the present application of a power demand forecasting method based on a deep fusion network includes a training phase S100 and a forecasting phase S200. The training phase S100 includes: S110, acquiring raw electricity consumption data and raw influencing factor data; S120, performing data cleaning, downsampling, normalization, and dataset partitioning on the raw electricity consumption data and raw influencing factor data to obtain a normalized training set and a normalized validation set; S130, screening key influencing factors on the normalized training set to obtain a set of significant influencing factors; S140, performing time lag correction and coupling relationship analysis on the set of significant influencing factors based on the electricity consumption data in the normalized training set to obtain a final feature set; S150, training the deep fusion network based on the final feature set and the normalized validation set to obtain a trained deep fusion network. The prediction phase S200 includes: S210, acquiring electricity consumption data and influencing factor data; S220, extracting the feature set to be predicted from the electricity consumption data and influencing factor data; S230, inputting the feature set to be predicted into the trained deep fusion network to obtain the predicted electricity demand value.

[0042] Specifically, in step S110, raw electricity consumption data and raw influencing factor data are acquired. It is understood that, because the construction of electricity demand forecasting models highly depends on the regularity of historical load changes and the driving effect of external environmental factors on electricity consumption behavior, a single data source is often insufficient to comprehensively depict the complex operating state of the power system. Furthermore, without high-quality initial input, subsequent feature mining and model training will be impossible. Therefore, in the technical solution of this application, raw electricity consumption data and raw influencing factor data are acquired to construct an initial data foundation that includes load time-series characteristics and multi-dimensional environmental characteristics. This ensures that subsequent data processing and model training have sufficient and comprehensive information input, thereby providing solid data support for uncovering the deep logic of electricity demand changes.

[0043] More specifically, in a concrete example of this application, historical load sequences and related environmental parameters of the target area are collected by establishing data transmission interfaces with the power dispatch automation system and external information service platforms. For the raw electricity consumption data, user electricity load records within a specified time span are exported by accessing the power grid company's smart metering system or data acquisition and monitoring system at a preset time granularity. These data records contain precise timestamps and corresponding active power values. The sampling frequency is set to once every thirty minutes to ensure that fine-grained fluctuations in daily load can be captured. Specifically, individual user data undergoes de-identification and aggregation processing, retaining only statistics relevant to prediction; data transmission uses TLS / HTTPS, and sensitive fields are stored with anonymized data. For the original influencing factor data, the historical meteorological database interface released by the meteorological bureau is called to obtain meteorological information such as temperature, relative humidity, rainfall and wind speed for the corresponding region during the same period. At the same time, macroeconomic and social activity indicators, including the GDP index, holiday schedule, electricity price adjustment policy release time and the date of major social events, are extracted from government statistical departments or relevant industry databases. The above-mentioned multi-source heterogeneous data covering meteorology, economy, policy and social activities are aggregated and stored in a unified database to prepare for the subsequent data cleaning and feature engineering stages.

[0044] Specifically, in step S120, the original electricity consumption data and original influencing factor data are cleaned, downsampled, normalized, and partitioned to obtain a normalized training set and a normalized validation set. It is understandable that due to the complex operating environment of the power acquisition terminal and the uncertainty of signal transmission, the acquired raw data often contains missing values ​​and abnormal noise. Furthermore, the electricity load values ​​differ from those of meteorological and economic influencing factors in terms of physical dimensions and magnitude. Directly inputting such multi-source, heterogeneous, and noisy data into a deep neural network can lead to model convergence difficulties or getting trapped in local optima. Simultaneously, the redundancy of high-frequency sampled data increases computational overhead and masks long-term trends. Therefore, in the technical solution of this application, the original electricity consumption data and original influencing factor data are further cleaned, downsampled, normalized, and partitioned to obtain a normalized training set and a normalized validation set. This removes invalid interference from the data, aggregates high-frequency instantaneous features into daily granular features that better reflect periodic patterns, eliminates the dimensional influence between different features, and constructs an independent sample set that conforms to temporal logic for model learning and parameter tuning. This effectively improves the quality and consistency of the input data, accelerates the gradient descent convergence process of the subsequent deep fusion network, and ensures the objectivity and generalization ability of the model evaluation results.

[0045] Figure 3This is a flowchart illustrating the process of data cleaning, downsampling, normalization, and dataset partitioning to obtain a normalized training set and a normalized validation set for a power demand forecasting method based on deep fusion networks, according to an embodiment of this application. Figure 3 As shown, step S120 includes: S121, cleaning the original electricity consumption data and the original influencing factor data to obtain cleaned electricity data and cleaned influencing factor data; S122, performing time-series aggregation of high-frequency electricity data on the cleaned electricity data to obtain daily electricity consumption data; S123, unifying the dimensions of the daily electricity consumption data and the cleaned influencing factor data across the entire feature space to obtain a normalized comprehensive dataset; S124, performing time-series logic-based dataset partitioning on the normalized comprehensive dataset to obtain a normalized training set and a normalized validation set.

[0046] In step S121, the original electricity consumption data and original influencing factor data are cleaned to obtain cleaned electricity data and cleaned influencing factor data. It is understood that due to equipment failures or network transmission delays in electricity metering devices, and environmental interference with meteorological monitoring equipment, the collected raw data often contains time discontinuities or abnormal noise points deviating from the normal physical range. These data quality issues directly disrupt the continuity of the time series and introduce biases, thereby misleading the model's feature learning direction. Therefore, in the technical solution of this application, the original electricity consumption data and original influencing factor data are further cleaned to obtain cleaned electricity data and cleaned influencing factor data. This fills in the gaps in the data sequence and corrects or removes unreasonable abnormal values. This restores the complete temporal logic of the data, eliminates the negative impact of noise on prediction accuracy, and provides an accurate and reliable data foundation for subsequent feature aggregation and model training.

[0047] More specifically, in a concrete example of this application, the collected raw electricity consumption data and raw influencing factor data are first scanned in their entirety to locate missing and outlier values. For the identified missing data points, temporal proximity interpolation is used for filling, that is, linear interpolation is calculated using adjacent valid data points before and after the missing time and their time distance to estimate the value at the missing time, ensuring a smooth transition of data over time. For the identified outlier data points, a reasonable threshold range is set based on the statistical distribution characteristics of historical data. For extreme values ​​exceeding this range, truncation is used to limit them to the threshold boundary, or a seasonal median replacement strategy is adopted, that is, the median of the historical data for the same period of the season to which the outlier belongs is selected to replace the outlier. This removes abrupt noise while preserving the inherent pattern of electricity load variation with the seasons, ultimately outputting cleaned electricity data and cleaned influencing factor data that are continuous and accurate due to the repair of missing and outlier values.

[0048] In step S122, the cleaned power data undergoes time-series aggregation of high-frequency power data to obtain daily electricity consumption data. It is understandable that while the original high-frequency sampling data records instantaneous fluctuations in grid load, the numerous micro-random disturbances it contains often obscure the macro-level changes in electricity demand on a daily scale. Furthermore, excessively high time resolution can lead to a sharp expansion of the feature dimensions of subsequent input models, increasing computational load and potentially introducing unnecessary noise interference. Considering that many external influencing factors are essentially statistically analyzed on a daily basis, the data granularity not only needs to match but also needs to focus on the total characteristics of the grid's daily dispatch plan. Therefore, in the technical solution of this application, the cleaned power data undergoes further time-series aggregation of high-frequency power data to obtain daily electricity consumption data. This transforms the fine-grained instantaneous load curve integral into a coarse-grained daily total load sequence, reducing data dimensionality while retaining the total daily electricity consumption information, thus achieving alignment of power load data with daily-dimensional influencing factor data at the time granularity. This effectively smooths out high-frequency noise, highlights the periodic patterns of electricity demand changes with seasons and date types, and provides a more robust and physically meaningful daily load characteristic input for deeply integrated networks.

[0049] More specifically, in a concrete example of this application, a time-window-based cumulative aggregation operation is performed. First, the number of data points per day corresponding to the sampling frequency is determined. For the cleaned power data with a sampling interval of thirty minutes, it is determined that each day contains forty-eight discrete sampling points. Then, the long-cycle power time-series data is segmented and sliced ​​according to the time boundary of the natural day. For the time window of the i-th natural day, the instantaneous power consumption values ​​of all thirty-minute intervals contained in the window are traversed, and the data of these forty-eight sampling points are accumulated and summed to obtain the total power consumption value for that day. This calculation process traverses the entire historical data period, and the accumulated results of each day are rearranged according to the date order, finally generating a daily power consumption data sequence consisting of a series of daily total power consumption values, completing the dimensional reduction from high-frequency instantaneous power to daily power consumption characteristics.

[0050] In step S123, the daily electricity consumption data and the cleaned influencing factor data are subjected to dimensional unification across the entire feature space to obtain a normalized comprehensive dataset. It is understandable that since daily electricity consumption data typically represents large values ​​of thousands or even tens of thousands of kilowatt-hours, while the numerical range of influencing factor data such as temperature, relative humidity, and economic indices is often smaller, directly inputting these differences in dimensions and magnitudes into the model without processing can lead to excessive weight dependence on large-value features during gradient descent in deep neural networks, thereby suppressing the expression of small-value features and even causing gradient oscillations that hinder convergence. Therefore, in the technical solution of this application, the daily electricity consumption data and the cleaned influencing factor data are further subjected to dimensional unification across the entire feature space to obtain a normalized comprehensive dataset. This eliminates scale interference caused by different physical dimensions, linearly mapping the feature values ​​of all dimensions to the same standard interval. This ensures that each feature has an equal initial status in model training, effectively improving the numerical stability and convergence speed of the optimization algorithm, and providing a uniformly distributed, high-quality input for the deep fusion network.

[0051] More specifically, in a concrete example of this application, a max-min normalization method is used to perform a unified linear transformation on multi-source data. First, the daily electricity consumption data aggregated over time is concatenated with the cleaned multi-dimensional influencing factor data to construct an original data matrix with a full feature space. Then, for each feature dimension in this matrix, the entire time series is traversed to identify and extract the global maximum and global minimum values ​​under that dimension. Based on this, a linear normalization formula is applied to subtract the minimum value of the corresponding dimension from each original data point in the matrix, and then divide by the range of that dimension. This compresses and maps all physical quantities that were originally distributed on different orders of magnitude into a closed interval between 0 and 1, ultimately generating a normalized comprehensive dataset that eliminates dimensional differences and preserves the original distribution pattern.

[0052] In step S124, the normalized comprehensive dataset is partitioned based on time-series logic to obtain a normalized training set and a normalized validation set. It is understood that, due to the time-series dependence and seasonal cyclical characteristics of electricity load data, using a conventional random shuffling method to partition the dataset would not only disrupt the inherent temporal continuity of the data but also potentially leak samples containing future information into the training process, resulting in inflated model evaluation results that fail to reflect true predictive ability. Furthermore, to effectively monitor overfitting and dynamically adjust hyperparameters during model training, an independent validation data subset needs to be constructed. Therefore, in the technical solution of this application, the normalized comprehensive dataset is further partitioned based on time-series logic to obtain a normalized training set and a normalized validation set. This constructs non-overlapping and functionally independent sample subsets for model parameter learning and performance verification while strictly preserving the chronological order. This ensures that the model infers future trends based solely on historical information, eliminates look-ahead bias, and improves the generalization performance and robustness of the deep fusion network through the feedback mechanism of the validation set.

[0053] More specifically, in a concrete example of this application, a time-indexed hierarchical slicing operation is performed. First, the normalized aggregate dataset is ensured to be arranged strictly in chronological order of date and time, prohibiting any form of random rearrangement. Then, splitting nodes are calculated based on the total sample size of the dataset, and a two-level partitioning strategy is used to construct the final set. In the first-level partition, the first 90% of the time series data is designated as model building data, and the remaining 10% is reserved for final testing. In the second-level partition, the model building data is further subdivided internally. The first 80% of the time series segments within this subset are extracted as a normalized training set to drive weight updates and feature learning in the deep neural network, while the last 20% of the time series segments are extracted as a normalized validation set to evaluate the model state and implement early stopping strategies during training iterations. Through the rigorous division based on the above time-series logic, the amount of data in the normalized training set is precisely 72% of the total amount of the normalized comprehensive dataset, while the amount of data in the normalized validation set is 18% of the total amount of the normalized comprehensive dataset. This ensures sufficient training data while providing statistically significant validation samples for model tuning.

[0054] Specifically, in step S130, the normalized training set is screened for key influencing factors to obtain a set of significant influencing factors. It is understandable that, due to the complexity and diversity of external factors affecting electricity demand, encompassing multiple dimensions such as meteorology, economics, and social activities, and the varying correlation patterns between these factors and electricity load—including simple linear correlations and complex nonlinear relationships such as the U-shaped distribution between temperature and load—retaining all features without screening or using only a single evaluation index with fixed weights often leads to data redundancy, and may even result in the omission of key features due to the dilution effect of linear and nonlinear indices. Therefore, in the technical solution of this application, the normalized training set is further screened for key influencing factors to obtain a set of significant influencing factors. This allows the construction of an adaptive evaluation system that is compatible with both linear and nonlinear relationships. Through multi-dimensional statistical analysis, the actual driving force of each potential factor on electricity load is accurately quantified, thereby eliminating irrelevant or weakly correlated redundant variables. This reduces the feature dimension and noise interference of the model input, ensuring that the selected feature set can reflect both obvious linear trends and capture implicit nonlinear fluctuations, providing a data foundation with strong interpretability and high confidence for the efficient training and high-precision prediction of subsequent deep fusion networks.

[0055] Figure 4 This is a flowchart illustrating the process of screening key influencing factors in a normalized training set to obtain a set of significant influencing factors, according to an embodiment of this application, of a power demand forecasting method based on a deep fusion network. Figure 4 As shown, step S130 includes: S131, extracting multi-dimensional statistical correlation features from the target vector and influencing factors in the normalized training set to obtain a correlation coefficient vector set; S132, performing comprehensive correlation coefficient fusion based on adaptive weighting of relational complexity on each correlation coefficient vector in the correlation coefficient vector set to obtain a comprehensive correlation coefficient set; S133, based on the comprehensive correlation coefficient set, performing threshold-based key feature screening on the original factor list to obtain a set of significant influencing factors.

[0056] In step S131, multi-dimensional statistical correlation features are extracted from the target vector and influencing factors in the normalized training set to obtain a correlation coefficient vector set. It is understandable that the coupling mechanism between electricity load and external influencing factors is extremely complex, exhibiting both simple linear correlations and rich nonlinear, monotonic, and complex functional dependencies. For example, industrial output and electricity consumption may show a linear positive correlation, while temperature and load often exhibit a nonlinear U-shaped curve characteristic. Relying solely on a single statistical indicator for evaluation will inevitably lead to bias, thus missing key factors with weak linear correlations but strong nonlinear driving forces. Therefore, in the technical solution of this application, multi-dimensional statistical correlation features are further extracted from the target vector and influencing factors in the normalized training set to obtain a correlation coefficient vector set. This is used to construct a multivariate evaluation matrix covering linear, rank, and generalized functional relationships, quantitatively extracting the correlation strength of each candidate option from different statistical dimensions. This comprehensively captures the diverse driving mechanisms of potential factors on electricity demand, avoiding information omissions due to methodological limitations, and providing rigorous and complete data support for the subsequent construction of comprehensive indicators.

[0057] More specifically, in a concrete example of this application, a parallel computing architecture is used to traverse each column of potential influencing factor data and the target electricity consumption vector in the normalized training set, and four differentiated statistical algorithms are applied to perform feature extraction. First, the Pearson correlation coefficient is calculated, which accurately measures the linear fit between influencing factors and electricity load by analyzing the ratio of covariance to standard deviation between variables. Second, the Kendall rank correlation coefficient is calculated, which assesses the strength of the ordered association between the two by statistically analyzing the proportion of consistent and inconsistent pairs between the influencing factor sequence and the load sequence based on rank consistency. Simultaneously, the Spearman rank correlation coefficient is calculated, which effectively assesses the monotonic nonlinear relationship between variables by converting the original data into rank and calculating its linear correlation. Finally, the maximum information coefficient is calculated based on mutual information theory. By discretizing the scatter plot of factors and load into grids of different resolutions, a grid partitioning method that maximizes mutual information is searched to capture any type of functional relationship, including complex periodic oscillations. For each influencing factor, the four statistical values ​​calculated above are encapsulated into an independent vector, and finally converged to form a correlation coefficient vector set describing the multidimensional correlation characteristics of the factor.

[0058] In step S132, the correlation coefficient vectors in the correlation coefficient vector set are fused using a comprehensive correlation coefficient fusion method based on adaptive weighting of relational complexity to obtain a comprehensive correlation coefficient set. It is understandable that, since the existing methods use an indiscriminate arithmetic mean fusion strategy to construct the comprehensive correlation coefficient, it is assumed that the four correlation measurement methods contribute equally to the final comprehensive correlation, i.e., the weights are fixed. However, these four methods differ fundamentally in their focus on capturing the relationships between variables: the Pearson correlation coefficient strictly measures linear relationships, the Kendall and Spearman correlation coefficients emphasize monotonic relationships, while the maximum information coefficient aims to capture a wide range of functional relationships, especially adept at discovering complex nonlinear relationships. In the actual scenario of electricity demand forecasting, the relationship between influencing factors and electricity consumption is not static. For example, the relationship between temperature and electricity consumption usually exhibits a U-shaped nonlinear characteristic, while the relationship between industrial output and electricity consumption may be closer to linear. If the Pearson coefficient, which focuses on linearity, and the MIC coefficient, which focuses on nonlinearity, are given equal importance, the high value of the MIC coefficient will be diluted by the low value of the Pearson coefficient in strongly nonlinear scenarios, resulting in the comprehensive index failing to accurately reflect the true strong correlation; conversely, the same applies in strongly linear scenarios. Therefore, in the technical solution of this application, the comprehensive correlation coefficient set is obtained by further fusing the various correlation coefficient vectors in the correlation coefficient vector set based on adaptive weighting of relational complexity. This introduces an index that can dynamically quantify relational complexity, adaptively assigning weights to linear and nonlinear correlation measures, thereby intelligently determining whether to focus on capturing complex nonlinear patterns such as temperature changes or stable linear trends such as economic growth based on data characteristics. In this way, the fundamental weakness of lacking adaptability to relational complexity can be effectively overcome, achieving more intelligent and accurate feature selection.

[0059] Figure 5 This is a flowchart illustrating a method for predicting electricity demand based on a deep fusion network, according to an embodiment of this application, which involves fusing comprehensive correlation coefficients from various correlation coefficient vectors in a correlation coefficient vector set using adaptive weighting based on relational complexity to obtain a comprehensive correlation coefficient set. For example... Figure 5 As shown, step S132 further includes: S1321, performing relational complexity quantification on the correlation coefficient vector to obtain a relational complexity index; S1322, determining adaptive weights based on the relational complexity index; and S1323, performing adaptive weighted fusion on the correlation coefficient vector based on the adaptive weights to obtain a comprehensive correlation coefficient.

[0060] In step S1321, the correlation coefficient vector is quantified to obtain a relationship complexity index. It is understandable that, due to the high heterogeneity of the coupling mechanisms between different influencing factors and power load, a single statistical index is insufficient to simultaneously encompass two drastically different modes: simple linear driving and complex nonlinear driving. The lack of an objective quantitative standard to determine the intrinsic attributes of data relationships directly hinders the accurate allocation of subsequent feature weights. Therefore, in the technical solution of this application, the correlation coefficient vector is further quantified to obtain a relationship complexity index, thereby establishing a data-driven objective standard to distinguish whether the dominant mode of the relationship between influencing factors and power demand is linear or nonlinear. This produces a quantitative value that accurately characterizes the relationship complexity, providing a decision-making basis for subsequent dynamic weight allocation and ensuring that the feature selection process adapts to the true patterns of data changes.

[0061] More specifically, in a concrete example of this application, a feature analysis process based on differences in statistical indicators is performed. Specifically, a relationship complexity index is constructed by comparing the MIC (Mutual Integrity Scale), which best represents a non-linear relationship, with the Pearson coefficient, which best represents a linear relationship. The core function of this indicator is to quantify the degree of non-linearity in the relationship between current influencing factors and electricity demand, expressed as:

[0062]

[0063] in, The relation complexity metric is a non-negative real number used to quantify the non-linear components in a relation. The maximum information coefficient measures the strength of any functional relationship between two variables. The Pearson correlation coefficient measures the strength of the linear relationship between two variables. This indicates the absolute value operation. This indicates taking the larger value between 0 and the calculated result to ensure the non-negativity of the indicator. Based on this calculation logic, it should be noted that when the MIC value is higher than the Pearson coefficient, it indicates a strong nonlinear relationship between the variables. For example, when analyzing the relationship between temperature and summer electricity consumption, the data usually exhibits a U-shaped nonlinear characteristic: the load fluctuates less within the human comfort temperature range (such as 22℃ to 26℃), but once the temperature exceeds the 30℃ high-temperature threshold, the cooling load will increase exponentially. At this time, the linear Pearson coefficient is often low, while the MIC coefficient is high, leading to a U-shaped nonlinear relationship. The values ​​will be larger, thus more sensitively capturing the nonlinear driving characteristics of meteorological factors; conversely, if the relationship leans towards linearity, the two values ​​will be close. The value tends to be close to 0. For example, when analyzing the relationship between industrial output and electricity consumption, as factory capacity increases, electricity consumption usually shows a stable, proportional growth trend. At this time, both the Pearson coefficient and the MIC coefficient are at high levels and have similar values, resulting in extremely low index values, accurately reflecting the linear driving nature of economic indicators. In this way, a quantitative value that can accurately characterize the complexity of the relationship is produced, providing a decision-making basis for subsequent dynamic allocation of weights.

[0064] In step S1322, adaptive weights are determined based on the relation complexity index. It is understandable that, since the relation complexity index calculated in the preceding steps is essentially a raw value reflecting the degree of difference between linear and nonlinear relationships, its numerical range is not a normalized interval, making it difficult to directly use as a weighting coefficient in the subsequent weighted summation formula. Furthermore, a nonlinear mapping mechanism is needed to achieve a sensitive response to different relational patterns. Therefore, in the technical solution of this application, adaptive weights are further determined based on the relation complexity index, thereby transforming the abstract complexity index into a concrete control signal that can be used for weighted fusion. This generates a weight signal that accurately reflects the relational pattern, thereby achieving differentiated processing of different types of correlation measurements and ensuring a smooth and intelligent switch between linear and nonlinear dominance in the feature fusion process.

[0065] More specifically, in a concrete example of this application, an adaptive fusion weight generation step is performed based on the complexity index obtained from the preceding steps. Specifically, the Sigmoid function is used to generate the relational complexity index. Mapped to a smooth, dynamically fused weight between 0 and 1. Its specific mathematical expression is shown in the following formula:

[0066]

[0067] in, For adaptive weights, a real number in the interval (0,1) is used. is the base of the natural logarithm. This is the gain coefficient, used to adjust the steepness of the weight curve, i.e., the sensitivity of the weights to changes in complexity. The center offset is used to set the response center point for the complexity index. This formula, used in this step, constructs an intelligent soft switch. In practical applications, when the relationship is highly non-linear (…), When dealing with large-scale (e.g., in the case of nonlinear exponential growth between air temperature and air conditioning power load during high-temperature periods in summer), this function will drive... A value close to 1 indicates that the mechanism should place greater trust in nonlinear correlation indicators, thus giving higher decision-making power to indicators such as the maximum information coefficient; conversely, when the linearity of the relationship is high ( When the time frame is small, for example when analyzing the linear relationship between basic electricity costs and fixed capacity, this function will drive... A value close to 0 means that linear correlation indicators should be trusted more, thus making full use of the high sensitivity of the Pearson coefficient in linearity measurement.

[0068] In step S1323, the correlation coefficient vector is adaptively weighted and fused to obtain the comprehensive correlation coefficient based on adaptive weights. It is understandable that existing methods employ an indiscriminate arithmetic average fusion strategy for constructing the comprehensive correlation coefficient, assuming that the four correlation measurement methods contribute equally to the final comprehensive correlation. This leads to unavoidable information dilution and evaluation bias issues caused by fixed-weight fusion when facing complex and variable power load influencing factors; that is, linear indicators may mask the value of nonlinear indicators, and vice versa. Therefore, in the technical solution of this application, the correlation coefficient vector is further adaptively weighted and fused to obtain the comprehensive correlation coefficient based on adaptive weights, thereby abandoning the simple arithmetic average and adopting a method based on adaptive weights. A precisely controlled weighted fusion strategy is employed to achieve intelligent fusion. The ultimate technical effect is the generation of a more accurate, robust, and comprehensive correlation coefficient that reflects the inherent relationship patterns within the data. This improvement not only enhances the accuracy of key influencing factor selection but also provides higher-quality feature inputs for downstream electricity demand forecasting models, potentially fundamentally improving the forecasting accuracy and decision support capabilities of the entire forecasting system.

[0069] More specifically, in a particular example of this application, an adaptive weighted fusion step is performed. Specifically, the Pearson coefficient is used as a representative of linear relationships, and the average of the other three coefficients (Kendall, Spearman, and maximum information coefficient) is used as a comprehensive representative of nonlinear and monotonic relationships. A dynamic weighted sum is then performed to generate an improved comprehensive correlation coefficient. The specific calculation logic is shown in the following formula:

[0070]

[0071] in, To improve the overall correlation coefficient, Kendall's correlation coefficient measures ordered association. The Spearman correlation coefficient measures rank association. Based on this calculation logic, it should be noted that the calculation of the comprehensive correlation coefficient no longer treats all metrics equally. In practical applications, if the relationship is linear ( For example, when analyzing the relationship between total electricity consumption and regional industrial added value, since the two often exhibit a highly positive linear correlation, the weight allocation in the formula will make... The value will be mainly determined by This decision allows for full utilization of the Pearson coefficient's advantage in linear measurement; conversely, if the relationship is highly nonlinear ( For example, when analyzing the relationship between residential electricity consumption and real-time temperature, due to the existence of a U-shaped nonlinear characteristic, at this time... The value will be primarily determined by the mean of three other coefficients that are better suited to measuring nonlinear or monotonic relationships, thus avoiding the masking effect of linear indicators on strong correlations. This allows the final composite correlation coefficient to intelligently adapt to the inherent patterns in the data, thereby more accurately assessing the importance of influencing factors and outputting a more confident and insightful correlation evaluation value.

[0072] In step S133, based on the comprehensive correlation coefficient set, the original factor list is subjected to threshold-based key feature screening to obtain a set of significant influencing factors. It is understandable that the initially collected list of influencing factors often covers extremely broad dimensions such as meteorology, economy, policy, and social activities, inevitably containing a large number of redundant features that contribute little or no correlation to changes in power load. Directly inputting this low-value, noisy data into subsequent computationally intensive deep learning models would not only increase computational resource consumption but also easily lead to the curse of dimensionality and model overfitting, thereby reducing the generalization ability of the prediction results. Therefore, in the technical solution of this application, the original factor list is further subjected to threshold-based key feature screening based on the comprehensive correlation coefficient set to obtain a set of significant influencing factors. This constructs a strict feature admission barrier, physically isolating high-driving factors from low-correlation noise based on quantitative evaluation values. This ensures that only key variables with predictive value enter the subsequent time-delay modeling and neural network training stages, thereby significantly improving the operating efficiency of the prediction system and the expression accuracy of core features while simplifying the model input structure.

[0073] More specifically, in a concrete example of this application, a rigorous numerical filtering operation is performed. First, a discrimination threshold is set based on the load characteristics of the power system and historical experience, for example, a threshold of 0.8. Then, the improved composite correlation coefficient value corresponding to each influencing factor in the composite correlation coefficient set is iterated and compared one by one with the preset threshold. During this process, factors with composite correlation coefficient values ​​greater than 0.8, such as real-time summer temperatures or output indices of major industrial sectors, are determined to have a strong correlation with electricity demand, are marked as significant influencing factors, and retained in the significant influencing factor set. Factors with composite correlation coefficient values ​​less than or equal to 0.8, such as a certain type of foreign exchange rate with low correlation to the local electricity consumption structure or micro-meteorological indicators in non-sensitive seasons, are determined to be weakly correlated or irrelevant features and are directly removed from the original factor list. Finally, a set of significant influencing factors containing only high-confidence features is output, providing accurate and focused input variables for subsequent time-lag effect analysis.

[0074] Specifically, in step S140, based on the electricity consumption data in the normalized training set, time-lag correction and coupling relationship analysis are performed on the set of significant influencing factors to obtain the final feature set. It is understandable that changes in electricity load are often not an immediate response to current external stimuli, but rather a lagged effect of the cumulative effects of factors such as meteorology, economics, and social activities over a period of time. Furthermore, there are often complex coupling interactions between different influencing factors; for example, high temperature and high humidity often work together to cause a surge in cooling load. Simply using current values ​​or ignoring the information overlap between factors will make it difficult for the model to capture the dynamic transmission mechanism and introduce a large number of redundant features. Therefore, in the technical solution of this application, time-lag correction and coupling relationship analysis are further performed on the set of significant influencing factors based on the electricity consumption data in the normalized training set to obtain the final feature set. This allows for precise quantification of the specific contribution weights of each key factor at different lag periods through mathematical modeling, and the identification and fusion of feature variables with strong synergistic change relationships. In this way, high-confidence input characteristics that conform to the physical operation laws and economic transmission logic of the power system can be effectively reconstructed. While eliminating multicollinearity in the feature space, the time-series cumulative effects and deep coupling interaction information that are crucial for load forecasting can be fully preserved.

[0075] More specifically, in a specific example of this application, based on electricity consumption data in the normalized training set, time lag correction and coupling relationship analysis are performed on the set of significant influencing factors to obtain the final feature set. This includes: constructing and estimating the parameters of an Almon multinomial distributed lag model on the electricity consumption data and the set of significant influencing factors in the normalized training set to obtain estimated polynomial coefficients; calculating the time lag contribution and dynamically correcting the features of the set of significant influencing factors based on the estimated polynomial coefficients to obtain corrected values ​​for the influencing factors; and performing coupling coordination degree analysis and fusion among the features of the set of significant influencing factors based on the corrected values ​​for the influencing factors to obtain the final feature set.

[0076] Accordingly, an Almon multinomial distributed lag model is constructed and its parameters estimated for the electricity consumption data and the set of significant influencing factors in the normalized training set to obtain estimated polynomial coefficients. It is understandable that the impact of external factors on electricity load often has time lags and cumulative effects. For example, rising temperatures do not immediately trigger peak air conditioning loads but rather exhibit a dynamic transmission process lasting several days. Directly introducing a large number of original variables with lag periods into the model would not only cause a surge in the number of parameters but also lead to severe multicollinearity problems, resulting in model estimation failure. Therefore, in the technical solution of this application, an Almon multinomial distributed lag model is further constructed and its parameters estimated for the electricity consumption data and the set of significant influencing factors in the normalized training set to obtain estimated polynomial coefficients. This allows for dimensionality reduction of the complex lag distribution structure through polynomial approximation, significantly reducing the number of parameters to be estimated while preserving complete time-series impact information. This effectively solves the multicollinearity problem and accurately captures the dynamic driving patterns of electricity demand under different time delays for each key factor.

[0077] More specifically, in a concrete example of this application, a regression analysis method based on the Almon transform is used to construct the model. First, for each specific feature in the set of significant influencing factors, such as daily maximum temperature or a macroeconomic index, a finite distribution lag model is constructed, with current normalized electricity consumption as the dependent variable and current and several past period values ​​of that feature as explanatory variables. Based on this, the Almon polynomial assumption is introduced, which assumes that the influence coefficients of each lag period do not change independently, but rather follow a low-order polynomial distribution with respect to the number of lag periods, typically set to a quadratic or cubic order. Subsequently, according to this polynomial constraint, the lagged variables in the original model are linearly reorganized and transformed, constructing a smaller number of new transformed variables, thus transforming the original model containing numerous lagged terms into a new linear regression model containing only a small number of polynomial coefficients. Finally, the least squares method is used to fit the transformed model to the training set data, and the equation is solved by minimizing the sum of squared residuals, thereby accurately obtaining the polynomial coefficients used to describe the estimated shape of the lag distribution.

[0078] Accordingly, based on the estimated polynomial coefficients, the time-delay contribution of the set of significant influencing factors is calculated and the features are dynamically corrected to obtain the corrected values ​​of the influencing factors. It is understandable that, since the estimated polynomial coefficients obtained from the Almon transform are merely intermediate mathematical parameters describing the lag distribution, they do not directly and intuitively reflect the physical influence weight of influencing factors on the current power load at a specific historical moment. Furthermore, if the original sequences of multiple lag periods are still retained as input, the problems of high feature dimensionality and redundancy in time series information cannot be completely solved. Therefore, in the technical solution of this application, the time-delay contribution of the set of significant influencing factors is further calculated and the features are dynamically corrected based on the estimated polynomial coefficients to obtain the corrected values ​​of the influencing factors. This allows for the reverse reconstruction of the specific influence distribution of each lag period from the polynomial space, and the compression of multi-period historical information into a single comprehensive feature value through weighted aggregation. In this way, the dynamic cumulative effect of external factors such as accumulated temperature or economic transmission on power demand can be accurately quantified, providing input features for the deep fusion network that combine temporal depth and dimensional simplicity.

[0079] More specifically, in a concrete example of this application, the inverse transformation of the Almon polynomial is first performed. Using the estimated polynomial coefficients output from regression analysis, combined with a preset polynomial order and maximum lag period, the original lag coefficients corresponding to each specific lag period are calculated period by period according to the polynomial definition formula. These lag coefficients physically represent the marginal effect of a unit change in the factor at a certain point in the past on the current load. Subsequently, contribution normalization calculation is performed. By extracting the absolute value of the lag coefficients for each period and calculating its proportion in the sum of the absolute values ​​of all lag coefficients, a standardized time-lag contribution index is generated. Finally, dynamic correction processing of the eigenvalues ​​is performed. For each time point in the concentration of significant influencing factors, a sliding window is used to obtain the historical data sequence within its corresponding lag period. The time-lag contribution obtained above is used as a weight to perform a weighted summation of these historical values. This process generates a sequence of corrected influencing factor values ​​that integrates historical cumulative effects, thereby transforming data such as temperature, which was originally discretely distributed over multiple days, into an effective cumulative temperature eigenvalue that can directly drive changes in the current load, achieving dynamic correction and time-series empowerment of the original physical quantity.

[0080] Accordingly, based on the corrected values ​​of influencing factors, the coupling coordination degree between features of the significant influencing factor set is analyzed and fused to obtain the final feature set. It is understandable that changes in electricity demand are not driven by a single isolated factor, but often stem from the synergistic interaction of multiple dimensions of factors such as meteorology, economics, and social activities. For example, high temperatures are often accompanied by high humidity, jointly increasing cooling load. If this nonlinear coupling effect between factors is ignored, it will not only make it difficult for the model to capture the load mutation patterns under the combined effect of multiple factors, but also cause information redundancy due to highly correlated feature inputs, reducing training efficiency. Therefore, in the technical solution of this application, the coupling coordination degree between features of the significant influencing factor set is further analyzed and fused based on the corrected values ​​of influencing factors to obtain the final feature set. This allows for the accurate assessment of the interaction strength and coordination level between various factors through a quantitative mathematical model, and the organic fusion and dimensionality reduction of features with strong coupling and synergistic relationships. This effectively eliminates redundant information in the feature space, extracts composite driving factors that better reflect the essence of the power system's operating state, thereby improving the input quality and prediction accuracy of the deep fusion network.

[0081] More specifically, in a concrete example of this application, a coupling coordination degree model from physics is used to conduct in-depth analysis of multidimensional feature data after time-delay correction. First, to eliminate differences in dimensions and orders of magnitude between the correction values ​​of different influencing factors, range standardization is performed on all features to be analyzed to ensure data comparability. Then, for any two feature vectors in the set of significant influencing factors, such as the corrected daily average temperature and daily average relative humidity, the coupling degree index is calculated using the coupling degree formula. This index reflects the strength of the interaction between the two systems. Based on this, combined with preset weighting coefficients, a comprehensive coordination index reflecting the overall development level of the two factors is calculated, and the final coupling coordination degree value is further obtained using the geometric mean method. For the calculated coupling coordination degree, a judgment threshold is set. For feature combinations with a coupling coordination degree higher than this threshold, a strong coupling synergy is determined, indicating that the two factors have a high degree of consistency and synergy in driving changes in power load. Then, methods such as weighted average or principal component analysis are used to merge these two features into a new comprehensive composite feature; while features with a coupling coordination degree lower than the threshold retain their independence. Finally, the retained independent features and the fused composite features are reassembled to construct a final feature set that is both concise and rich in interactive information, which serves as the input vector for subsequent deep fusion network models.

[0082] Specifically, in step S150, the deep fusion network is trained based on the final feature set and the normalized validation set to obtain a trained deep fusion network. It is understandable that, because power load data exhibits both local fluctuations influenced by sudden environmental factors and long-term trends dominated by seasonal cycles, a single network structure is insufficient to effectively model these two distinct dynamic patterns simultaneously. Furthermore, without a rigorous validation mechanism during training, the model is highly susceptible to overfitting, leading to a significant decrease in generalization performance on unknown data. Therefore, in the technical solution of this application, the deep fusion network is further trained based on the final feature set and the normalized validation set to obtain a trained deep fusion network. This drives the model to learn the nonlinear mapping relationship between multi-source coupling features and power demand through an iterative optimization algorithm, and the validation set is used to monitor the training status in real time to adjust hyperparameters. This effectively captures the multi-scale temporal dependencies of the load, avoids the gradient vanishing problem, and ultimately generates a prediction model with strong robustness and high accuracy.

[0083] More specifically, in a concrete example of this application, a hybrid architecture integrating a convolutional neural network and a bidirectional long short-term memory network is first constructed. A one-dimensional convolutional neural network branch is configured to extract multi-scale local temporal features from the final input feature set through convolution operations and the SELU activation function. Simultaneously, a bidirectional long short-term memory network branch is configured to capture long-term bidirectional temporal dependencies from both forward and backward dimensions. Based on this, a deep fusion training process is executed. The final feature set from the training set is input into the model, and the intermediate feature representations of the two branches are fused element-wise at multiple depth levels of the network. The fused features are then passed to the next layer through residual connections to ensure information flow and gradient stability in the deep network. Subsequently, the Adam optimizer (…) is employed. =0.9, The parameters are iterated using a set value of 0.999. SELU activation is used for the CNN branch, and the bidirectional LSTM uses tanh and sigmoid gating. LayerNorm is added before the fully connected layer to stabilize training. The gradient is calculated based on the error between the model output value and the true value, and the convolutional kernel weights, gating unit parameters, and bias terms in the network are dynamically updated. In each iteration, a normalized validation set is periodically input into the current model for forward propagation evaluation. The error change curve of the model on the validation set is monitored in real time. When the validation error no longer decreases or shows an upward trend, an early stopping mechanism is triggered, thereby selecting and saving the model parameter combination that performs best on the validation set, and outputting the final trained deep fusion network. In an optional implementation, the one-dimensional convolutional neural network branch includes three convolutional layers. The number of convolutional kernels in the first, second, and third convolutional layers are 32, 64, and 64, respectively, and the kernel size is 3 with a stride of 1. The bidirectional long short-term memory network branch includes two bidirectional long short-term memory layers, each with 64 hidden units, and a dropout layer with a dropout rate of 0.2 is set between each layer. During the training phase, mean squared error is used as the loss function, the batch size is set to 64, and the initial learning rate is [missing information]. The maximum number of learning rounds is 100, and the actual number of training rounds is automatically determined through the aforementioned early stop mechanism.

[0084] Specifically, in step S210, electricity consumption data and influencing factor data are acquired. It is understood that since the inference process of a deep learning model is essentially based on the mapping calculation of the target variable between known historical states and future boundary conditions, and the operation of a power system has inertial characteristics and external stimulus response characteristics, if the latest load trajectory up to the current moment and the external environmental parameters for the period to be predicted are lacking, the trained model will be unable to initiate forward propagation calculations due to the lack of input basis. Therefore, in the technical solution of this application, electricity consumption data and influencing factor data are acquired to construct a complete spatiotemporal context input for the prediction engine, including the current operating state of the system and future external driving conditions. This ensures that the deep fusion network has data support consistent with the structure of the training phase during the prediction phase, thereby generating accurate future electricity demand predictions based on the latest real-time information.

[0085] More specifically, in a concrete example of this application, real-time data stream acquisition and integration operations are performed. For electricity consumption data, the real-time database of the power dispatch automation system or electricity information collection system is accessed through a secure interface to extract the historical electricity load time series up to the current moment. This series must cover the shortest historical time window required to satisfy the Almon multinomial lag model to ensure that complete time lag characteristics can be calculated. For influencing factor data, the numerical weather prediction interface of the meteorological service provider is accessed to obtain meteorological forecast data such as temperature, humidity, and precipitation probability for the forecast date. At the same time, historical meteorological observation records from a past period are retrieved to match the time lag calculation requirements. In addition, the latest published values ​​of holiday arrangements, workday identifiers, and macroeconomic indicators are read synchronously from the calendar database. The real-time historical data and future forecast data from different sources are timestamped and cached to construct the raw input data stream used to drive the prediction model.

[0086] In step S220, a feature set to be predicted is extracted from electricity consumption data and influencing factor data. It is understood that because the raw data obtained in the prediction stage differs from the input tensors required by the trained deep fusion network in terms of data structure, physical dimensions, and feature distribution, and because the trained model is based on specific hysteresis correction features and coupled collaborative features for parameter optimization, directly inputting unprocessed raw values ​​will lead to dimensionality mismatch or feature semantic misalignment, failing to activate the nonlinear mapping relationship learned by the model. Therefore, in the technical solution of this application, a feature set to be predicted is further extracted from electricity consumption data and influencing factor data. This transforms and reconstructs the real-time multi-source raw data stream strictly according to the feature engineering logic determined in the training stage, generating a high-dimensional feature matrix consistent with the spatial distribution of the training samples. This ensures that the data input to the model maintains a high degree of isomorphism with the training data in terms of statistical characteristics and physical meaning, thereby guaranteeing the prediction accuracy and stability of the deep fusion network in the inference stage.

[0087] More specifically, in a concrete example of this application, a feature reconstruction process based on model metadata is executed. First, the data processing parameters saved during the training phase are invoked, including the maximum and minimum values ​​required for normalization, a list of significant influencing factors, and the Almon polynomial lag coefficients and coupling model thresholds. Next, the real-time acquired electricity and influencing factor data are cleaned and aggregated at the daily granularity, and a linear transformation is performed using the saved normalization parameters to map the new data to the same numerical range as the training set. Subsequently, key variables are selected from the full dataset based on the list of significant influencing factors. Using the saved lag model parameters combined with historical data windows up to the current time, the time-lag correction values ​​for each factor are calculated, and feature pairs with strong coupling relationships are calculated and fused. Finally, the processed independent and coupled features are vectorized and assembled according to the order required by the model input layer to construct the final feature set to be predicted.

[0088] In step S230, the feature set to be predicted is input into the trained deep fusion network to obtain the predicted electricity demand value. It is understood that, since the trained deep fusion network has internalized the complex nonlinear mapping relationship and high-dimensional spatiotemporal dependency logic between multi-source heterogeneous features and electricity demand into the network's weights and bias parameters, simply constructing the feature set is insufficient to directly obtain the prediction result; the model's inference logic must be executed to activate these parameters' response to new data. Therefore, in the technical solution of this application, the feature set to be predicted is further input into the trained deep fusion network to obtain the predicted electricity demand value, thereby triggering the forward propagation mechanism of the deep neural network. The fixed model parameters are used to perform layer-by-layer abstraction and regression calculations on the newly input feature data. This outputs an accurate future electricity demand value that integrates historical time lag effects, multi-factor coupling effects, and long- and short-term time series dependencies, providing reliable data support for grid dispatching decisions.

[0089] More specifically, in a concrete example of this application, the inference engine first loads the network structure definition and trained weight file of the deep fusion network, and feeds the feature set to be predicted, formatted as a time-series tensor, into the input layer of the model. Inside the model, a one-dimensional convolutional neural network branch performs convolution operations on the input tensor to capture the local variation trend and multi-scale morphological features of the load curve. Simultaneously, a bidirectional long short-term memory network branch processes the sequence along the time step to capture bidirectional time dependencies. The feature vectors output by these two branches are summed element-wise and nonlinearly transformed in the deep fusion layer, and then passed to the fully connected regression layer at the end of the network through residual connections. The regression layer outputs a normalized prediction value between zero and one. Finally, using the historical maximum and minimum electricity consumption parameters recorded during the data preparation phase, an inverse normalization operation is performed on the normalized prediction value to map it back to the actual physical dimensions of electricity consumption, thereby generating the final interpretable electricity demand prediction value.

[0090] In summary, a power demand forecasting method based on a deep fusion network, according to embodiments of this application, is explained. First, it performs standardized preprocessing on multi-source heterogeneous power and influencing factor data. Addressing the technical problem in multi-dimensional feature selection where fixed weights prevent the consideration of both linear and nonlinear relationships, an adaptive weighting mechanism based on relational complexity is introduced to dynamically fuse multiple statistical correlation coefficients to accurately identify significant influencing factors. Subsequently, using the Almon multinomial distributed lag model and coupling coordination analysis, the time-lag contribution of influencing factors is quantified, and the coupling interaction between features is handled, constructing a final feature set containing a dynamic transmission mechanism. Finally, by constructing a deep fusion model integrating a convolutional neural network and a bidirectional long short-term memory network, multi-scale local features and bidirectional long-term time-series dependencies are extracted in parallel. High-precision power demand forecasting is achieved through deep fusion with residual connections, overcoming the shortcomings of traditional models such as weak generalization ability and difficulty in capturing complex time-series patterns.

[0091] Furthermore, a power demand forecasting system based on a deep fusion network is also provided.

[0092] Figure 6 This is a block diagram of a power demand forecasting system based on a deep fusion network according to an embodiment of this application. Figure 6 As shown, an embodiment of this application discloses a power demand forecasting system 100 based on a deep fusion network, comprising a training system 110 and a forecasting system 120. The training system 110 includes: a raw data acquisition module 111 for acquiring raw power consumption data and raw influencing factor data; a normalized training set and normalized validation set acquisition module 112 for performing data cleaning, data downsampling, data normalization, and dataset partitioning on the raw power consumption data and raw influencing factor data to obtain a normalized training set and a normalized validation set; a key influencing factor screening module 113 for screening key influencing factors on the normalized training set to obtain a set of significant influencing factors; a time lag correction and coupling relationship analysis module 114 for performing time lag correction and coupling relationship analysis on the set of significant influencing factors based on the power consumption data in the normalized training set to obtain a final feature set; and a deep fusion network training module 115 for training the deep fusion network based on the final feature set and the normalized validation set to obtain a trained deep fusion network. The prediction system 120 includes: a data acquisition module 121 for acquiring electricity consumption data and influencing factor data; a data extraction module 122 for extracting a feature set to be predicted from the electricity consumption data and influencing factor data; and an electricity demand prediction value acquisition module 123 for inputting the feature set to be predicted into a trained deep fusion network to obtain the electricity demand prediction value.

[0093] As described above, a power demand forecasting system 100 based on a deep fusion network according to an embodiment of this application can be implemented in various computing devices, such as servers deployed in power dispatch centers, power grid energy management system workstations, or edge computing terminals of smart grids. In one possible implementation, the power demand forecasting system 100 based on a deep fusion network according to an embodiment of this application can be integrated into the computing device as a software module and / or hardware module. For example, the power demand forecasting system 100 based on a deep fusion network can be a functional component in the dispatch automation system of the computing device, or it can be a dedicated load forecasting application developed for the computing device; of course, these functional logics can also be embedded in the hardware of the computing device, for example, implemented as dedicated graphics processing unit logic or field-programmable gate arrays to achieve efficient real-time forecasting.

[0094] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A power demand prediction method based on a deep fusion network, characterized in that, The method comprises: a training stage and a prediction stage; wherein the training stage comprises the steps of: obtaining original power consumption data and original influencing factor data; performing data cleaning, data downsampling, data normalization and data set division on the original power consumption data and the original influencing factor data to obtain a normalized training set and a normalized verification set; performing key influencing factor screening on the normalized training set to obtain a significant influencing factor set; based on the power consumption data in the normalized training set, performing time lag correction and coupling relationship analysis on the significant influencing factor set to obtain a final feature set; training a deep fusion network based on the final feature set and the normalized verification set to obtain a trained deep fusion network; wherein the prediction stage comprises the steps of: obtaining power consumption data and influencing factor data; extracting a to-be-predicted feature set from the power consumption data and the influencing factor data; inputting the to-be-predicted feature set into the trained deep fusion network to obtain a power demand prediction value. 2.The deep fusion network based power demand forecasting method of claim 1, wherein, The data cleaning, data downsampling, data normalization and data set division on the original power consumption data and the original influencing factor data to obtain the normalized training set and the normalized verification set comprise: performing data cleaning on the original power consumption data and the original influencing factor data to obtain cleaned power data and cleaned influencing factor data; performing time series aggregation of high-frequency power data on the cleaned power data to obtain daily power consumption data; performing dimension unification in the whole feature space on the daily power consumption data and the cleaned influencing factor data to obtain a normalized comprehensive data set; performing time series logic-based data set division on the normalized comprehensive data set to obtain the normalized training set and the normalized verification set. 3.The deep fusion network based power demand forecasting method of claim 2, wherein, The normalized training set accounts for 72% of the normalized comprehensive data set, and the normalized verification set accounts for 18% of the normalized comprehensive data set. 4.The deep fusion network based power demand forecasting method of claim 1, wherein, The key influencing factor screening on the normalized training set to obtain the significant influencing factor set comprises: performing multi-dimensional statistical correlation feature extraction on target vectors and influencing factors in the normalized training set to obtain a correlation coefficient vector set; performing comprehensive correlation coefficient fusion based on adaptive weighting of relationship complexity on each correlation coefficient vector in the correlation coefficient vector set to obtain a comprehensive correlation coefficient set; based on the comprehensive correlation coefficient set, performing key feature screening based on a threshold on an original factor list to obtain the significant influencing factor set. 5.The deep fusion network based power demand forecasting method of claim 4, wherein, The comprehensive correlation coefficient fusion based on adaptive weighting of relationship complexity on each correlation coefficient vector in the correlation coefficient vector set to obtain a comprehensive correlation coefficient set comprises: quantifying the relationship complexity of the correlation coefficient vector to obtain a relationship complexity index; determining an adaptive weight based on the relationship complexity index; based on the adaptive weight, adaptively weighting and fusing the correlation coefficient vector to obtain a comprehensive correlation coefficient. 6.The deep fusion network based power demand forecasting method of claim 5, wherein, The quantification of the relationship complexity of the correlation coefficient vector to obtain a relationship complexity index comprises: quantifying the relationship complexity of the correlation coefficient vector according to the following formula: wherein, is a relationship complexity indicator, is a maximum information coefficient, is a Pearson correlation coefficient, is a representation of the absolute value operation. is a representation of the maximum of 0 and the result of the calculation. 7.The deep fusion network based power demand forecasting method of claim 1, wherein, The time lag correction and coupling relationship analysis on the significant influencing factor set based on the power consumption data in the normalized training set to obtain a final feature set comprises: The Almon polynomial distribution lag model is constructed and parameter estimation is performed on the power consumption data in the normalized training set and the set of significant influencing factors to obtain estimated polynomial coefficients; Based on the estimated polynomial coefficients, time lag contribution calculation and feature dynamic correction are performed on the set of significant influencing factors to obtain influencing factor correction values; Based on the influencing factor correction values, coupling coordination degree analysis and fusion are performed among the set of significant influencing factors to obtain the final feature set. 8.A power demand prediction system based on deep fusion network, characterized in that, It comprises: a training system and a prediction system; wherein the training system comprises: an original data acquisition module for acquiring original power consumption data and original influencing factor data; a normalized training set and a normalized validation set acquisition module for performing data cleaning, data downsampling, data normalization and data set division on the original power consumption data and the original influencing factor data to obtain a normalized training set and a normalized validation set; a key influencing factor screening module for screening key influencing factors from the normalized training set to obtain a set of significant influencing factors; a time lag correction and coupling relationship analysis module for performing time lag correction and coupling relationship analysis on the set of significant influencing factors based on the power consumption data in the normalized training set to obtain a final feature set; a deep fusion network training module for training a deep fusion network based on the final feature set and the normalized validation set to obtain a trained deep fusion network; wherein the prediction system comprises: a data acquisition module for acquiring power consumption data and influencing factor data; a data extraction module for extracting a to-be-predicted feature set from the power consumption data and the influencing factor data; a power demand prediction value acquisition module for inputting the to-be-predicted feature set into the trained deep fusion network to obtain a power demand prediction value. 9.The power demand prediction system based on deep fusion network according to claim 8, wherein, The key influencing factor screening module comprises: a multi-dimensional statistical correlation feature extraction unit for performing multi-dimensional statistical correlation feature extraction on target vectors and influencing factors in the normalized training set to obtain a set of correlation coefficient vectors; a comprehensive correlation coefficient fusion unit for performing comprehensive correlation coefficient fusion on each correlation coefficient vector in the set of correlation coefficient vectors based on relationship complexity adaptive weighting to obtain a set of comprehensive correlation coefficients; a key feature screening unit for performing key feature screening on an original factor list based on a threshold based on the set of comprehensive correlation coefficients to obtain a set of significant influencing factors. 10.The power demand prediction system based on deep fusion network according to claim 8, wherein, The time lag correction and coupling relationship analysis module comprises: a model construction and parameter estimation unit for constructing and parameter estimation of the Almon polynomial distribution lag model on the power consumption data in the normalized training set and the set of significant influencing factors to obtain estimated polynomial coefficients; a time lag contribution calculation and feature dynamic correction unit for performing time lag contribution calculation and feature dynamic correction on the set of significant influencing factors based on the estimated polynomial coefficients to obtain influencing factor correction values; a coupling coordination degree analysis and fusion unit for performing coupling coordination degree analysis and fusion among the set of significant influencing factors based on the influencing factor correction values to obtain the final feature set.

Images