A power grid-oriented electricity sales volume prediction and deviation closed-loop correction method and device

Abstract

The present application relates to a kind of power grid-oriented electricity sales forecast and deviation closed-loop correction method and device, belong to power system operation side data analysis and artificial intelligence application technical field.The method includes: first, the multi-source feature of fusing meteorology, industry economy and power grid business disturbance is constructed, and baseline prediction is generated using the adaptive baseline decomposition model of time-varying weight gatekeeping.Then, by graph neural network (GNN) modeling " industry-region " spatial correlation, and using Transformer encoder captures multivariate nonlinear coupling characteristics, outputs fine prediction.Finally, introduce the closed-loop feedback mechanism based on deep residual strategy gradient, utilize GRU network memory history error evolution inertia, drive strategy network to generate dynamic correction, to integrated open-loop prediction result is compensated in real time.The method significantly improves the prediction accuracy and robustness in complex scenarios, effectively reduces the deviation assessment risk in power spot market.

Description

Technical Field

[0001] This invention belongs to the field of power system operation-side data analysis and artificial intelligence application technology, specifically relating to a method and device for predicting electricity sales and correcting deviations in power grids. Background Technology

[0002] In existing power system electricity sales forecasting technologies, a common approach employs linear time series extrapolation based on traditional statistics. This method primarily relies on mining the autocorrelation and periodicity of historical electricity data to construct a parameterized linear regression model for future predictions. This approach typically assumes that the data generation process follows a stationary or weakly stationary distribution, focusing on capturing the linear trend and seasonal components of the time series. While this method offers advantages such as low computational complexity and strong interpretability in scenarios with stable loads or single trends, its expressive power is limited by the endogeneity of the linear assumption. When faced with the impact of extreme weather (e.g., unstructured weather type coding), sudden business disruptions (e.g., a surge in fault repair orders), and the coupling of multi-source heterogeneous data caused by macroeconomic restructuring, the model struggles to construct effective nonlinear mapping relationships. Furthermore, this approach lacks a feature fusion mechanism for multidimensional external variables. When dealing with complex calendar effects such as Spring Festival holiday adjustments, it cannot dynamically compensate by introducing external covariates, leading to a significant decrease in prediction accuracy in atypical scenarios.

[0003] Another existing approach is based on end-to-end deep learning fitting techniques. It utilizes deep neural network architectures such as Long Short-Term Memory (LSTM) networks or Temporal Convolutional Networks (TCNs) to learn a high-dimensional nonlinear mapping between input features and target electricity sales volume through supervised training on a large number of historical samples. For example, it directly trains a regression network using historical meteorological data and historical load sequences to capture complex fluctuation patterns. While this type of method can solve the problem of multivariate coupled modeling to some extent and significantly improve the fitting accuracy of nonlinear features, it typically employs an "open-loop" prediction mode. This means that model parameters remain frozen after offline training, lacking a feedback correction mechanism for real-time error distribution. When electricity spot market rules are fine-tuned or data distribution undergoes conceptual drift (e.g., systematic deviations in weather forecasts or structural changes in load components), the model cannot adaptively adjust its prediction strategy using recent historical deviation information. Furthermore, this type of method often behaves as a "black box" system lacking physical meaning, lacking decoupled modeling of long-term trend terms and short-term fluctuation terms, making it difficult to meet the business needs of power grid operation and management for interpretability of prediction results and manual verification.

[0004] Another similar approach is the independent forecasting method based on single-point time series. Most existing electricity sales forecasting technologies (whether statistical or deep learning-based) treat each region or industry as an independent forecasting unit, using only its own historical data for modeling. Their main drawback is that they ignore the objectively existing "spatial-industry transmission effect" on the power grid side. For example, fluctuations in electricity consumption in upstream raw material industries often transmit to downstream manufacturing through the supply chain, and there is significant synergy in economic activities between adjacent administrative regions. Existing methods artificially sever this spatial and logical horizontal connection, causing the model to become an "information silo." When faced with missing local data or sudden anomalies at a single node, it cannot utilize information from surrounding nodes or related industries for auxiliary inference, severely limiting the robustness of the forecasting system. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for predicting electricity sales and correcting deviations in power grids. It is no longer limited to single-point time series analysis, but first constructs a multi-source heterogeneous feature engineering that covers refined meteorology, industry-specific macroeconomic indices and fault reporting business. It also innovatively introduces spatial graph neural network (GNN) to perform topological modeling of the economic transmission relationship between "industry-region", thereby effectively breaking the limitation of "information silos" in single-point prediction.

[0006] In the model prediction process, this invention employs a hierarchical decoupling and dynamic fusion strategy: On the one hand, it designs an adaptive baseline decomposition mechanism based on time-varying weight gating, utilizing local fluctuation entropy to dynamically adjust the weight combination of trend, seasonal, and holiday components, rather than using a fixed additive decomposition. This ensures that the model strictly adheres to strong business rules such as the Spring Festival cliff, while significantly improving its adaptability to non-steady-state data. On the other hand, it combines a Transformer deep attention network to capture the high-dimensional nonlinear coupling characteristics between multi-source variables. Furthermore, to address the concept drift that may be caused by adjustments to spot market rules or sudden disturbances (such as extreme weather forecast deviations), this invention introduces a closed-loop feedback mechanism based on deep residual policy gradients. Through the GRU network, it senses the evolutionary inertia of historical deviations and guides the policy network to generate dynamic correction compensation, achieving real-time offsetting of systematic errors in the model. Through these methods, this invention can simultaneously consider business interpretability, spatial synergy, and environmental adaptability in power grid operation-side forecasting, effectively overcoming the problems of difficulty in multi-source data fusion, weak anti-interference ability, and lack of adaptive correction capability in existing technologies.

[0007] To achieve the above objectives, the technical solution of the present invention is: a method for predicting electricity sales and correcting deviations in a power grid, comprising:

[0008] We acquire refined meteorological characteristics, industry-specific macroeconomic characteristics, and power grid business interference characteristics from the power grid operation side, and perform spatiotemporal alignment processing to construct a comprehensive feature vector at time t. ;

[0009] Based on historical electricity sales data Baseline predictions for time t are generated through adaptive baseline decomposition. ;

[0010] Construct a graph structure with administrative power supply bureaus and key industries as nodes and economic dependencies as edges. V and E represent the sets of nodes and edges, respectively, and the feature vectors of each node are... Mapping to a graph structure, spatial correlation features are extracted using graph convolutional networks. ;

[0011] Integrate spatial correlation features The historical feature sequence is input into the Transformer encoder, which uses a multi-head self-attention mechanism for nonlinear coupled feature extraction and outputs refined prediction values. ;

[0012] The meta-learner is trained based on the validation set data, generating dynamic weight coefficients. For baseline predicted values Compared with detailed forecast values We perform weighted ensemble to obtain open-loop prediction results. ;

[0013] Construct a multidimensional deviation state vector based on historical prediction deviation sequences. The error hidden state vector at time t is obtained by encoding using a gated recurrent unit (GRU) network. The error hidden state vector is input into the deep policy network to generate dynamic correction values. The dynamic correction is added to the open-loop prediction result to obtain the final electricity sales forecast. .

[0014] Furthermore, the refined meteorological characteristics include the daily maximum temperature. Daily minimum temperature and weather type feature vectors after unique thermal encoding The aforementioned industry-specific macroeconomic characteristics are daily industry prosperity index vectors obtained by interpolating monthly electricity sales data for each industry. The power grid service interference characteristics include the number of fault repair work orders. and average repair time .

[0015] Furthermore, the adaptive baseline decomposition includes the following steps:

[0016] Define the business baseline at time t The components are trend component g(t), seasonal component s(t), and holiday component. The weighted sum;

[0017] The trend component g(t) is defined by a piecewise linear growth model:

[0018]

[0019] in, Based on the growth rate, Basic intercept, Adjust the vector for the growth rate. As the time-varying point indicator vector, The intercept smoothing vector is used to maintain the continuity of the function;

[0020] The seasonal component s(t) is defined by a Fourier series:

[0021]

[0022] in, For the basic cycle, Let be the order of the Fourier series. These are the Fourier coefficients;

[0023] The holiday component Defined as:

[0024]

[0025] in, The impact matrix of holidays, This is a vector of the impact coefficients of holidays. With a mean of 0 and a variance of It follows a normal distribution.

[0026] Furthermore, the adaptive baseline decomposition also includes:

[0027] Calculate the electricity sales sequence within the time window Local information entropy within ;

[0028] Local information entropy Meteorological feature vector at the current moment By concatenating the features, we obtain the joint feature vector. ;

[0029] Joint feature vectors Input the gating network and calculate the time-varying weight coefficients:

[0030]

[0031] in, and For learnable gating parameters, The dynamic weights correspond to the trend component, seasonal component, and holiday component, respectively.

[0032] Furthermore, the baseline prediction value The calculation formula is:

[0033] .

[0034] Furthermore, the graph structure edge weight The construction methods include: constructing static weights based on the direct consumption coefficients in the industry input-output table, and / or constructing dynamic weights based on the Pearson correlation coefficients of the historical electricity sales sequence of each node, and setting a threshold for the correlation coefficients to truncate them in order to maintain the sparsity of the graph structure.

[0035] The feature propagation formula for the graph convolutional network is:

[0036]

[0037] in, To add a self-loop adjacency matrix, For degree matrix, For learnable weight matrix, It is a non-linear activation function. This represents the feature matrix of the nodes in the l-th layer of the graph convolutional network. = .

[0038] Furthermore, the open-loop prediction results The calculation formula is:

[0039] .

[0040] Furthermore, the multidimensional deviation state vector is calculated using the following formula:

[0041]

[0042] in This is the deviation from the previous day. This is a deviation from the same day last week. The mean deviation within the sliding window. The first difference is the deviation;

[0043] The error hidden state vector is calculated as follows:

[0044]

[0045] The formula for calculating the dynamic correction Δt generated by the deep policy network is as follows:

[0046]

[0047] in, Here are the weight matrix and bias vector of the hidden layer. The weight matrix and bias vector of the output layer. To correct the amplitude scaling factor.

[0048] The present invention also provides a power grid electricity sales forecasting and deviation closed-loop correction device, comprising:

[0049] The feature construction module acquires refined meteorological features, industry-specific macroeconomic features, and power grid business interference features from the power grid operation side, and performs spatiotemporal alignment processing to construct a comprehensive feature vector at time t. ;

[0050] Adaptive baseline decomposition module, based on historical electricity sales sequence Baseline predictions for time t are generated through adaptive baseline decomposition. ;

[0051] The spatial feature extraction module constructs a graph structure with administrative power supply bureaus and key industries as nodes and economic dependencies as edges. V and E represent the sets of nodes and edges, respectively, and the feature vectors of each node are... Mapping to a graph structure, spatial correlation features are extracted using graph convolutional networks. ;

[0052] The multi-source regression prediction module will integrate spatial correlation features. The historical feature sequence is input into the Transformer encoder, which uses a multi-head self-attention mechanism for nonlinear coupled feature extraction and outputs refined prediction values. ;

[0053] The integrated prediction module trains a meta-learner based on validation set data to generate dynamic weight coefficients. For baseline predicted values Compared with detailed forecast values We perform weighted ensemble to obtain open-loop prediction results. ;

[0054] The closed-loop feedback correction module constructs a multi-dimensional deviation state vector based on historical prediction deviation sequences. The error hidden state vector is obtained by encoding using a gated recurrent unit (GRU) network. The error hidden state vector is input into the deep policy network to generate dynamic correction values. The dynamic correction is added to the open-loop prediction result to obtain the final electricity sales forecast. .

[0055] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This invention constructs a multi-source feature space that integrates fault repair work orders, industry-specific business climate indices, and refined meteorological data. It innovatively utilizes cubic spline interpolation to address the frequency mismatch between monthly economic data and daily forecasts, and introduces repair business characteristics to explicitly quantify the "passive reduction" effect of unplanned outages on electricity sales, thus resolving the logical mismatch between operational-side electricity sales and physical load from the data source.

[0058] 2. This invention proposes a decoupled modeling strategy that combines "preservation of deterministic rules" with "adaptation to stochastic fluctuations." It utilizes Prophet to explicitly extract holidays and long-term trends, constructing an inviolable business baseline; simultaneously, it leverages the self-attention mechanism of Transformer to capture nonlinear residual coupling between multiple variables. This architecture effectively overcomes the "black box" limitation of single deep learning models, achieving stable adherence to strong rules (such as the Spring Festival) and agile adaptation to weak rules (such as weather fluctuations).

[0059] 3. This invention proposes a decoupled modeling strategy that combines "preservation of deterministic rules" with "adaptation to stochastic fluctuations." It utilizes Prophet to explicitly extract holidays and long-term trends, constructing an inviolable business baseline; simultaneously, it leverages the self-attention mechanism of Transformer to capture nonlinear residual coupling between multiple variables. This architecture effectively overcomes the "black box" limitation of single deep learning models, achieving stable adherence to strong rules (such as the Spring Festival) and agile adaptation to weak rules (such as weather fluctuations). Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0061] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0062] This invention provides a method for predicting electricity sales and correcting deviations in a closed-loop manner for power grid applications, comprising:

[0063] We acquire refined meteorological characteristics, industry-specific macroeconomic characteristics, and power grid business interference characteristics from the power grid operation side, and perform spatiotemporal alignment processing to construct a comprehensive feature vector at time t. ;

[0064] Based on historical electricity sales data Baseline predictions for time t are generated through adaptive baseline decomposition. ;

[0065] Construct a graph structure with administrative power supply bureaus and key industries as nodes and economic dependencies as edges. V and E represent the sets of nodes and edges, respectively, and the feature vectors of each node are... Mapping to a graph structure, spatial correlation features are extracted using graph convolutional networks. ;

[0066] Integrate spatial correlation features The historical feature sequence is input into the Transformer encoder, which uses a multi-head self-attention mechanism for nonlinear coupled feature extraction and outputs refined prediction values. ;

[0067] The meta-learner is trained based on the validation set data, generating dynamic weight coefficients. For baseline predicted values Compared with detailed forecast values We perform weighted ensemble to obtain open-loop prediction results. ;

[0068] Construct a multidimensional deviation state vector based on historical prediction deviation sequences. The error hidden state vector at time t is obtained by encoding using a gated recurrent unit (GRU) network. The error hidden state vector is input into the deep policy network to generate dynamic correction values. The dynamic correction is added to the open-loop prediction result to obtain the final electricity sales forecast. .

[0069] The present invention also provides a power grid electricity sales forecasting and deviation closed-loop correction device, comprising:

[0070] The feature construction module acquires refined meteorological features, industry-specific macroeconomic features, and power grid business interference features from the power grid operation side, and performs spatiotemporal alignment processing to construct a comprehensive feature vector at time t. ;

[0071] Adaptive baseline decomposition module, based on historical electricity sales sequence Baseline predictions for time t are generated through adaptive baseline decomposition. ;

[0072] The spatial feature extraction module constructs a graph structure with administrative power supply bureaus and key industries as nodes and economic dependencies as edges. V and E represent the sets of nodes and edges, respectively, and the feature vectors of each node are... Mapping to a graph structure, spatial correlation features are extracted using graph convolutional networks. ;

[0073] The multi-source regression prediction module will integrate spatial correlation features. The historical feature sequence is input into the Transformer encoder, which uses a multi-head self-attention mechanism for nonlinear coupled feature extraction and outputs refined prediction values. ;

[0074] The integrated prediction module trains a meta-learner based on validation set data to generate dynamic weight coefficients. For baseline predicted values Compared with detailed forecast values We perform weighted ensemble to obtain open-loop prediction results. ;

[0075] The closed-loop feedback correction module constructs a multi-dimensional deviation state vector based on historical prediction deviation sequences. The error hidden state vector is obtained by encoding using a gated recurrent unit (GRU) network. The error hidden state vector is input into the deep policy network to generate dynamic correction values. The dynamic correction is added to the open-loop prediction result to obtain the final electricity sales forecast. .

[0076] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0077] The following is a detailed implementation process of the present invention.

[0078] like Figure 1 As shown, this invention proposes a method for power grid electricity sales forecasting and deviation closed-loop correction. This method no longer relies solely on physical load patterns, but instead constructs a hybrid forecasting framework integrating multi-dimensional features of "meteorological environment, industry economics, and business behavior." It innovatively introduces spatial map correlation and deep strategy gradient feedback mechanisms to address the pain point of concept drift in forecasting models under spot market conditions. The specific implementation process of this invention mainly includes the following core steps:

[0079] 1. Construction and Spatiotemporal Alignment Feature Engineering of Multi-Source Heterogeneous Data

[0080] This invention first establishes a multi-dimensional feature engineering approach oriented towards settlement standards, aiming to capture the core factors affecting electricity sales from the data source. Input data feature vector. It is composed of the following three types of heterogeneous data after spatiotemporal alignment:

[0081] The first category is refined meteorological features. Unlike traditional methods that only use average temperature, this invention extracts the daily maximum temperature based on meteorological monitoring data. Daily minimum temperature Furthermore, one-hot encoding is used to vectorize non-numerical weather conditions (such as "heavy rain", "typhoon", and "cloudy"). Let the set of weather types be... Then the first Weather feature vector of the day This initiative aims to capture the non-linear impacts of extreme weather events on residents' lives and businesses.

[0082] The second category is industry-specific macroeconomic characteristics. Considering that industrial production constitutes the main component of electricity sales, and that different industries have varying sensitivities to economic cycles, this invention selects monthly electricity sales data from industries that significantly impact regional economies (based on actual data, such as textiles, electronics, real estate, and finance). To address the frequency mismatch between monthly data and daily forecasts, cubic spline interpolation or a smoothing ladder preservation method is used to map monthly industry indices to a daily scale, constructing an industry prosperity index vector. It is used to capture the long-term trend impact of changes in industrial structure.

[0083] The third category is power grid business interference characteristics. This is the key to distinguishing between "physical load" and "operational electricity sales." This invention introduces the daily number of fault repair work orders. and average repair time These two indicators directly reflect the "passive reduction" in electricity consumption caused by equipment failure or unplanned outages. The constructed comprehensive feature vector... This is the concatenation of the above vectors.

[0084] 2. Adaptive baseline decomposition based on time-varying weight gating

[0085] To enhance the model's adaptability to non-stationary data while preserving its interpretability, this invention departs from the traditional additive decomposition model with fixed parameters. Instead, it proposes an adaptive baseline decomposition mechanism based on time-varying weight gating. This mechanism dynamically adjusts the weight combination of trend, seasonal, and holiday terms according to the local fluctuation entropy values ​​of the electricity sales series.

[0086] Let the time series of electricity sales be: This invention treats the observed data as a linear superposition of deterministic regular components and random fluctuation components. To explicitly characterize the stability of this structure, this invention introduces an additive decomposition operator. The business baseline at time t Defined as:

[0087]

[0088] in Each represents a set of parameters for a given component.

[0089] For the trend component g(t), to address the problem that traditional linear regression cannot adapt to the cyclical changes in macroeconomic structures (such as the prosperity of the textile and electronics industries), this invention adopts a piecewise linear growth model. Let... Adjust the vector for the variable point rate. If the time-varying point indicator vector is used, then the trend term is defined as:

[0090]

[0091] The physical meanings of the mathematical symbols in the formula are as follows:

[0092] This represents the basic growth rate of electricity sales, reflecting the natural economic growth rate under the condition of no major external policy interference;

[0093] : Represents the basic intercept of the electricity sales series;

[0094] : indicates in The growth rate adjustment vector at each inflection point is used to capture trend changes caused by industrial policy adjustments or major economic events.

[0095] : is the time indicator vector, when time Exceeding the A turning point At that time, the corresponding element in the vector Otherwise, it is 0;

[0096] The intercept smoothing vector is introduced to maintain the continuity of the trend function at the changing point. The calculation method for each element is as follows: .

[0097] This formula introduces sparse variable point parameters, enabling the model to automatically "memorize" the turning points when an industry's economy shifts from prosperity to decline or from decline to prosperity, thus forming a stable tracking of long-term trends.

[0098] For the seasonal component s(t), to ensure that the model strictly follows the annual / weekly cycle of the physical world, this invention maps it to the Fourier frequency domain. Let P be the fundamental period (e.g., 365.25 days), and constrain the waveform by constructing an orthogonal basis function set:

[0099]

[0100] The physical meanings of the mathematical symbols in the formula are as follows:

[0101] : Represents the seasonal component of the electricity sales sequence, used to capture periodic fluctuations in the physical world that strictly follow an annual / weekly cycle.

[0102] : This represents the basic cycle (such as 365.25 days mentioned in the text), which corresponds to the time cycle scale in the physical world.

[0103] : Indicates the order (or number of harmonics) of the Fourier series, which determines the precision of the fitting of waveform details.

[0104] : Represents the Fourier coefficients, used to determine the amplitude (weight) of sine and cosine waves at different frequencies, thereby locking the fluctuation phase of electricity sales.

[0105] This indicator locks in the fluctuation phase of electricity sales in the frequency domain. When the input data experiences short-term distortion due to extreme weather, the baseline output can still remain within a reasonable physical phase range due to the global smoothing properties of the Fourier series, thus serving as a noise reduction anchor.

[0106] For the holiday component h(t), this invention introduces a set of indicator functions H for a specific time window (covering Spring Festival, National Day, and make-up workdays). The holiday influence matrix Z(t) is defined, where elements at non-holiday times are 0, then:

[0107]

[0108] The physical meanings of the mathematical symbols in the formula are as follows:

[0109] : Indicates the holiday component, used to reflect the impact of specific time windows such as Spring Festival, National Day, and make-up workdays on electricity sales.

[0110] : Represents the holiday impact matrix (or reduced-dimensional matrix), which is a set of indicator functions passed through a specific time window. Defined as an element that is 0 on non-holiday periods.

[0111] : Represents the various impact coefficients (weights) of holidays, which follow a normal distribution. This constraint restricts the prediction results to a logical framework that conforms to the rules of the Chinese calendar (such as predictions that defy common sense, like preventing a surge in electricity consumption during the Spring Festival).

[0112] This constraint restricts the forecasting process to a logical framework that conforms to the rules of the Chinese calendar. For example, the extent of the power outage during the Spring Festival is determined by parameters. Explicit control, by its very mechanism, eliminates predictions that defy common sense, such as "a surge in electricity consumption during the Spring Festival."

[0113] Through the above mechanism, this invention achieves endogenous protection of power grid business rules without the need for a large number of labeled anomalous samples, enabling the baseline prediction results to adaptively match the evolutionary structure of macroeconomic and calendar rules, thereby effectively reducing the risk of forgetting physical common sense in the subsequent multi-source fusion stage.

[0114] Unlike the simple linear superposition of baseline components in existing technologies, this invention introduces dynamic gating coefficients. To adapt to unsteady environments. First, calculate the time window. Local information entropy within the sequence In this embodiment, the sliding window size The preferred cycle is 7 or 14 days (corresponding to a weekly / bi-weekly cycle in the physical world). Specifically, sample entropy is used to quantify the disorder of the current electricity data. This is calculated as the negative logarithm of the probability of sequence pattern matching within a window; a larger value indicates more severe fluctuations. Subsequently, the calculated scalar... With meteorological feature vector Perform dimensional concatenation. Let... The concatenated joint feature vector Based on this joint feature, a gated network is constructed: in, For learnable gating parameters, The dynamic weights correspond to trends, seasons, and holidays, respectively. Through this design, the model can integrate the "chaos level of the sequence itself" with "external weather stimuli" to dynamically sense whether the system is in a period of severe fluctuation.

[0115] The final adaptive baseline prediction formula is:

[0116]

[0117] 3. Spatial Dimension Graph Neural Networks (GNNs)

[0118] To address the issue that existing power grid operation-side electricity sales forecasting technologies only focus on a single regional time dimension and ignore economic transmission links between industries or regions, this invention innovatively introduces a spatial graph neural network to perform topological modeling of implicit relationships between multidimensional entities. Firstly...

[0119] Construct a graph structure to describe the relationships between prediction units. ,in:

[0120] Node set : Represents different forecast entities, including power supply bureau nodes in different administrative regions, as well as key industry nodes such as textiles, electronics, and finance;

[0121] edge set : Represents the economic dependencies between nodes. Edge weight. The construction involves two parallel approaches: First, based on the "Industry Input-Output Tables" regularly published by national or provincial statistical bureaus, extract industry... With the industry The direct consumption coefficient between nodes is used as a fixed physical-economic edge weight; secondly, dynamic edges are constructed based on the Pearson correlation coefficient of the historical electricity sales sequence of each node. To ensure the sparsity of the graph and computational efficiency, a correlation coefficient threshold is set for truncation (e.g., selecting a correlation coefficient threshold). Alternatively, for each node, only the top-K neighboring nodes with the highest correlation coefficients can be retained. Typically, the weight is set to 3 to 5. Edge weights above the threshold are retained, otherwise they are set to 0.

[0122] Then convert the feature vectors of each node Mapping to a graph structure, spatial correlation features are extracted using a spectral domain graph convolutional network (GCN). The feature propagation formula for a layer is defined as follows:

[0123]

[0124] in: This is an adjacency matrix with self-loops added. It is a degree matrix; The weight matrix is ​​a learnable weight matrix; This is a non-linear activation function. The spatial features output in this step... It will be integrated with time-series features to enter the next stage.

[0125] 4. Multi-source refined regression based on Transformer

[0126] To address the complex nonlinear coupling between meteorological, economic, and operational data, and the spatial features extracted by GNNs, this invention constructs a deep prediction model based on a Transformer encoder. First, historical sequences are encoded at location or embedded temporally to preserve temporal information. Then, the model incorporates spatial features... historical sequence The input model dynamically calculates the weight distribution of different variables at different time points using a multi-head self-attention mechanism. The core calculation formula is as follows:

[0127]

[0128] The physical meanings of the mathematical symbols in the formula are as follows:

[0129] The query matrix represents the active feature vector that needs to "pay attention to other information" at the current time step, which is derived from the historical sequence input that incorporates spatial features.

[0130] The key matrix represents the feature identifiers of each time step in the historical sequence, used for... Perform matching to calculate relevance.

[0131] The value matrix represents the actual content information (such as weather and load values) contained in the historical sequence, which is aggregated based on attention weights.

[0132] : Represents the dimension of the key vector. Here, it serves as a scaling factor to prevent the dot product result from becoming too large, causing the Softmax function to enter a saturation region with minimal gradients, thus ensuring the stability of model training.

[0133] The normalized exponential function transforms the matching score into a probability distribution of weights, ensuring that the sum of the weights of all points of interest is 1.

[0134] Through a self-attention mechanism, the model can dynamically capture complex logic such as "high temperatures usually lead to increased load, but if accompanied by a large number of fault reports, actual electricity sales may decrease." After processing by a feedforward neural network (FFN), the output is a refined prediction value containing high-frequency fluctuation information. .

[0135] To combine the stability of baseline models with the agility of deep models, this invention utilizes a stacking ensemble learning strategy for model fusion. A meta-learner is trained based on validation set data, employing Ridge Regression and XGBoost to dynamically generate weight coefficients. .

[0136] The formula for calculating the open-loop prediction result is:

[0137]

[0138] This strategy allows the model to assign higher weights to the adaptive baseline model during stable periods, while assigning higher weights to the graph neural network and the Transformer during atypical fluctuation periods, thus achieving adaptive switching for different scenarios.

[0139] 5. Closed-loop feedback correction based on deep residual policy gradient

[0140] In view of the fact that existing feedback control technologies mostly use simple PID linear control or shallow MLP, which are difficult to deal with nonlinear and long-delay error evolution problems, this invention proposes an intelligent feedback mechanism based on deep residual strategy gradient.

[0141] For the current time t, the system backtracks the historical deviation sequence to construct a multidimensional deviation state vector. :

[0142]

[0143] in, Characterized by short-term memory (yesterday bias). Characterizes cyclical memory (same week, same day deviation). This represents the average deviation (drift trend) within the sliding window. This is the first-order difference of the deviation (deterioration rate). This vector In the state space, it can effectively reflect the current "health" of the model: when When the system fluctuates near the zero vector, it indicates that the system is in a steady state; when A significant directional shift indicates that the system has experienced conceptual drift. A gated recurrent unit (GRU) is then used to perform deep encoding on the sequence to capture the time dependence of the error (such as continuous positive bias and periodic deviation oscillations).

[0144]

[0145] in For a moment The error hidden state vector implies the "drift inertia" of the system at the current moment. : Indicates the error state of the input at the current moment, that is, the deviation characteristics between the predicted value and the true value.

[0146] Then, based on the dynamic adjustment of the policy network, the hidden state is generated. Input a deep policy network, which directly outputs the best correction policy at the current time step. :

[0147]

[0148] The physical meanings of the mathematical symbols in the formula are as follows:

[0149] : The scaling factor for the correction magnitude is used to restore the normalized output to the physical magnitude of the actual electricity sales volume, preventing the correction from being too large and causing the forecast value to "overshoot".

[0150] Function: Limits the output to a reasonable range to prevent overshoot correction.

[0151] : Represents the learnable weight matrix and bias vector of the hidden layer of a deep policy network, used to extract nonlinear correction features from high-dimensional error hidden states.

[0152] : Represents the weight matrix and bias vector of the output layer, which maps the extracted features to the final corrected scalar.

[0153] Modify the activation function of the linear unit to introduce nonlinear factors and enhance the network's ability to fit complex error patterns.

[0154] In this embodiment, the deep policy network employs a multilayer perceptron (MLP) structure. This network contains at least one hidden layer with 128 neurons, and uses the ReLU activation function (i.e., the one in the formula). and The corresponding layer); the output layer contains one neuron and uses the Tanh activation function to map the output to... Interval (i.e., in the formula) and (the corresponding layer), then multiplied by a scaling factor of physical magnitude. This structure achieves a nonlinear mapping from high-dimensional hidden states to a single scalar correction with minimal computational overhead.

[0155] To ensure the correct direction of the policy network's corrections, an auxiliary loss function is constructed. During training, the network parameters are jointly optimized using gradient descent.

[0156] Finally, the feedback correction is added to the open-loop forecast result to obtain the final predicted electricity sales volume on the operational side:

[0157]

[0158] 6. Core defect overcoming mechanism and model accuracy verification

[0159] This invention addresses the unique challenge of the intertwined nature of physical laws and business interference in electricity sales forecasting scenarios by constructing a unified framework with deep synergistic effects. Existing technologies (such as pure GNNs or pure deep learning) often compromise on one aspect while neglecting another. This invention dynamically decouples the statistical baseline from the deep model through "adaptive gating," and forcibly couples open-loop prediction with real-time settlement errors through "policy gradient feedback." This "decoupling before closing the loop" design allows the spatial features extracted by GNNs and the nonlinear features extracted by Transformers to function within a controlled physical baseline (Prophet+) framework, effectively avoiding the logical instability and overfitting that can result from relying solely on deep learning or graph networks.

[0160] First, to address the issue that single-point prediction models ignore the correlation between industries and regions, this invention introduces spatial graph neural networks (GNN) and multi-head attention mechanisms to construct a spatiotemporally coupled feature space that integrates "meteorological environment - industry economy - business behavior". Without destroying the physical meaning of the data, the fluctuation characteristics of adjacent regions and upstream and downstream industries are used to assist in the inference of the current node, which significantly improves the prediction robustness in data sparse scenarios.

[0161] Secondly, to address the problem that traditional baseline models struggle to cope with non-steady-state disturbances, this invention proposes an adaptive baseline decomposition mechanism based on time-varying weight gating. By dynamically adjusting the weight combination of trends and holidays using local information entropy, it achieves both explicit adherence to strong rules such as the Spring Festival cliff and agile adaptation to sudden oscillations.

[0162] Finally, addressing the pain point of concept drift inherent in "open-loop" models, this invention innovatively employs a closed-loop feedback mechanism based on deep residual policy gradients. This mechanism utilizes a GRU network to memorize error evolution inertia and generate a dynamic correction strategy, achieving real-time offsetting of systematic deviations caused by changes in spot market rules. The specific technical overcoming mechanism is as follows:

[0163] (1) Addressing the problem of poor adaptability of static baseline models: This invention overcomes the shortcomings of traditional Prophet models, which have fixed parameters and are difficult to cope with sudden oscillations, by using "adaptive baseline based on time-varying weight gating". The technical effect is that: by using information entropy to dynamically perceive data fluctuations, the trend term weight is automatically strengthened during the stable period to ensure long-term accuracy, and the holiday term weight is automatically strengthened during the oscillation period to adapt to short-term shocks, thus realizing the "follow-up" of the model strategy.

[0164] (2) Addressing the problem of "information silos" in single-point prediction: This invention overcomes the shortcomings of existing technologies that ignore the correlation between industries / regions by using "Spatial Graph Neural Network (GNN)". The technical effect is that it constructs an economic transmission map on the operational side, enabling the model to "leverage its strengths" and use the characteristics of upstream industries or neighboring regions to assist in the prediction of the current node, significantly improving the inference ability in data-sparse or abnormal scenarios.

[0165] (3) Addressing the issues of lag and lack of memory in feedback control: This invention overcomes the shortcomings of traditional PID or MLP feedback, which lack historical memory and can only perform static compensation, by using "deep residual strategy gradient feedback". The technical effect is that the GRU network remembers the evolutionary inertia of the error (such as the pattern of "overestimation after a typhoon"), realizing the leap from "post-event correction" to "pre-event prediction and compensation", minimizing the risk of deviation assessment.

[0166] Furthermore, experiments simulated the streaming data scenarios faced by the power grid operators in the real physical world. The model was required to adapt quickly to current weather and market changes while strictly adhering to historical business patterns as it experienced different seasonal climates (stable spring, high summer temperatures, and abnormal typhoons) and business scenarios (normal workdays and major holidays) in chronological order. Specifically, a test sequence was constructed using daily electricity sales data from a coastal provincial capital city for 12 consecutive months, focusing on four typical scenarios: S = {Stable transition period (March-April) → High summer temperatures (July-August) → Typhoon / fault-prone period (August-September) → Spring Festival holiday (January-February)}. Five typical prediction limit strategies were compared in the experiments.

[0167] (1) ARIMA-X: A traditional statistical method with external regression. It introduces temperature as a regression variable, but it does not have the ability to fit nonlinearly;

[0168] (2) Single-Node-DL: Single-node deep learning. It uses a long short-term memory network for end-to-end prediction, without baseline constraints and closed-loop feedback;

[0169] (3) Static-Decomp: Static decomposition model. It only uses trend and seasonal terms for decomposition and does not include multi-source feature fusion and incremental adaptation.

[0170] (4) XGBoost-Reg: A gradient-based decision tree ensemble method. Inputting weather, repair requests, and industry characteristics, it represents the strong fitting ability of traditional machine learning on tabular data, but lacks a temporal feedback mechanism;

[0171] (5) TCN-Net: Temporal Convolutional Network. It uses causal convolution to capture temporal features and represents a competitive approach to convolutional architectures in deep learning.

[0172] The experiment uses the hybrid closed-loop model of this invention as the core. On the one hand, it measures the model’s mean absolute percentage error (MAPE, the lower the value, the better) over the entire time period. On the other hand, it focuses on evaluating the model’s practical effect in dealing with the deviation assessment of the spot market (i.e., the number of days with deviation > 2%) through “monthly assessment risk days” (RD). RD is directly related to the operating penalty risk of power grid companies (the lower the value, the better).

[0173] Table 1 Comparison of Sales Volume Forecasting Performance and Assessment Risk in Different Scenarios

[0174] ARIMA-X 2.15% 4.80% 6.20% 5.50% 4.66% 8.5 Single-Node-DL 1.80% 2.10% 3.50% 1.90% 2.90% 4.2 Static-Decomp 2.30% 3.10% 4.50% 4.20% 2.95% 5.0 XGBoost-Reg 1.65% 1.85% 2.40% 3.80% 2.42% 3.2 TCN-Net 1.75% 1.95% 3.20% 4.05% 2.74% 3.8 This invention 1.20% 1.35% 1.85% 1.50% 1.47% 0.8

[0175] The experimental results are shown in Table 1. As can be seen from the data in Table 1, ARIMA-X suffered severe accuracy failure when handling nonlinear, strongly disturbed scenarios (summer high temperatures, typhoon failures), with a MAPE as high as 6.20% in stage S3, indicating that the linear model cannot capture the complex coupling between "temperature" and "fault". While XGBoost-Reg achieved better performance than statistical models (MAPE 1.65%) during the stable transition period (S1) due to its strong fitting ability to table features, its error soared to 3.80% in stage S4 (Spring Festival). This is because tree models inherently lack the ability to extrapolate time series trends and cannot "predict" structural changes during long holidays based on calendar rules, unlike Prophet. Static-Decomp and TCN-Net performed reasonably well during the high-temperature period (S2), but their errors rebounded significantly during the typhoon / fault period (S3) due to the lack of explicit input of repair data; and during the Spring Festival holiday (S4), they exhibited obvious overfitting due to the lack of rule constraints. Single-Node-DL performed excellently during the Spring Festival holiday, but its response was sluggish under high temperatures and abnormal weather conditions.

[0176] In comparison, this invention maintains optimal stability across all scenarios. Particularly during the typhoon / failure period (S3), which is the most challenging for the model, thanks to the multi-source attention mechanism's keen capture of meteorological types (such as "heavy rain") and operational data, the MAPE is controlled at 1.85%, significantly outperforming other baselines. More importantly, regarding the "monthly average assessment risk days" indicator, which is crucial for operational efficiency, this invention, through its bias closed-loop feedback mechanism, reduces the number of days potentially incurring fines per month from the industry average of 4-5 days to 0.8 days. This verifies the stabilizing effect of the invention's "baseline knowledge maintenance" on holidays and the suppressive effect of "closed-loop incremental adaptation" on concept drift. When a potential systematic bias in the model is detected during typhoon days, the feedback network can dynamically adjust the correction amount. This avoids the model's oscillation between "blind confidence" and "reactive lag," and truly achieves reliable prediction of electricity sales on the operational side.

[0177] 7. Real-world application and effectiveness verification in business scenarios

[0178] To further verify the feasibility and business value of this invention in a real power grid spot market environment, an online trial operation was conducted for 12 months based on the actual business system of a provincial power grid company. Specific implementation data and scenario descriptions are as follows:

[0179] 7.1 Description of Experimental Data Base and Operating Environment

[0180] This invention integrates the provincial network marketing system, PMS equipment asset system, and meteorological dedicated line data. The training set uses historical data from the past 36 months (approximately 1095 days), while the test set consists of streaming validation data from the most recent 12 consecutive months. In terms of feature engineering, the Spatial Topology Graph Network (GNN) constructs 24 core nodes, including spatial nodes for 14 municipal-level power supply bureaus and 10 nodes for high-energy-consuming and key industries. Daily input features include 96 points of high-frequency meteorological data (temperature, humidity, wind, pressure, and unique weather type codes), statistical features from over 5000 daily fault repair work orders, and monthly updated prosperity indices for 10 major industries. The aligned feature vector reaches 128 dimensions. Regarding operational performance, the model is deployed on an enterprise-level server equipped with dual NVIDIA A100 GPUs, based on the PyTorch deep learning framework. Actual business scheduling results show that the response time for a single closed-loop forecast of electricity sales for the next 7 days in 14 cities of a province is strictly controlled within 1.5 seconds, which fully meets the rolling reporting timeliness requirements of the power grid for the day-ahead market and intraday market.

[0181] 7.2 Micro-analysis of typical business scenarios

[0182] To visually demonstrate the actual synergistic effect of the "multi-source fusion" and "closed-loop correction" mechanisms in this invention, a real business segment from the test set, depicting a scenario of "extreme weather conditions caused by a strong typhoon combined with widespread equipment outages," is selected for illustration. In August of a certain year, a super typhoon made landfall in the province, causing widespread power grid outages in three coastal cities (nodes A, B, and C), resulting in a surge in repair requests. Simultaneously, an upstream core manufacturing plant (node ​​D) suspended operations due to typhoon emergency response. Faced with this complex scenario, traditional models such as TCN-Net or single-node LSTM could only detect the sudden drop in temperature, providing an open-loop prediction downward adjustment of approximately 3%, far less than the actual power consumption drop of up to 12%. In contrast, the GNN spatial synergy mechanism of this invention, through the historical correlation of industry node D, keenly captured the economic effects of the upstream shutdown transmitting to the coastal cities; simultaneously, the multi-head attention mechanism assigned extremely high weight to the surge in repair requests, enabling the open-loop baseline prediction value to be accurately adjusted downward by 8% beforehand. On the day following the typhoon's landfall, when the model detected a systematic negative bias of -3.5% in the previous day's settled electricity volume, the deep policy gradient mechanism quickly took effect. The GRU network encoded this "overestimation bias inertia" into a hidden state, guiding the policy network to output an additional -3.8% correction. Ultimately, this invention kept the prediction bias within 1.8% for three consecutive days after the typhoon's passage, perfectly avoiding the spot market's bias threshold of over 2%.

[0183] 7.3 Business Economic Benefit Evaluation

[0184] Under the settlement rules of the electricity spot market, if the prediction deviation rate exceeds the exemption range (usually ±2%), the electricity seller will face high deviation penalty fees. Based on the above 12 months of online backtesting data for a full-cycle evaluation, using the original static-decomp prediction method, the provincial power grid would experience approximately 60 days of over-limit assessment risk per year. After adopting the solution of this invention, thanks to the real-time prediction and dynamic offsetting of systematic errors by the strategy gradient network, the annual assessment risk days in the province have been sharply reduced to 9 days, an overall reduction of up to 85%. If we conservatively estimate based on the province's average daily electricity sales volume and the average deviation settlement price in the spot market (taking a penalty of 0.05 yuan / kWh as an example), this invention, by significantly improving the prediction accuracy in extreme scenarios and non-steady-state business periods, is expected to reduce deviation assessment costs for the provincial power grid and related electricity sellers by several million yuan annually, demonstrating extremely significant economic benefits and engineering promotion value.

[0185] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for predicting electricity sales and correcting deviations in power grids, characterized in that, include: We acquire refined meteorological characteristics, industry-specific macroeconomic characteristics, and power grid business interference characteristics from the power grid operation side, and perform spatiotemporal alignment processing to construct a comprehensive feature vector at time t. ; Based on historical electricity sales data Baseline predictions for time t are generated through adaptive baseline decomposition. ; Construct a graph structure with administrative power supply bureaus and key industries as nodes and economic dependencies as edges. V and E represent the sets of nodes and edges, respectively, and the feature vectors of each node are... Mapping to a graph structure, spatial correlation features are extracted using graph convolutional networks. ; Integrate spatial correlation features The historical feature sequence is input into the Transformer encoder, which uses a multi-head self-attention mechanism for nonlinear coupled feature extraction and outputs refined prediction values. ; The meta-learner is trained based on the validation set data, generating dynamic weight coefficients. For baseline predicted values Compared with detailed forecasts We perform weighted ensemble to obtain open-loop prediction results. ; Construct a multidimensional deviation state vector based on historical prediction deviation sequences. The error hidden state vector at time t is obtained by encoding using a gated recurrent unit (GRU) network. The error hidden state vector is input into the deep policy network to generate dynamic correction values. The dynamic correction is added to the open-loop prediction result to obtain the final electricity sales forecast. .

2. The method for predicting electricity sales and correcting deviations in a power grid according to claim 1, characterized in that, The refined meteorological characteristics include the daily maximum temperature. Daily minimum temperature and weather type feature vectors after unique thermal encoding The aforementioned industry-specific macroeconomic characteristics are daily industry prosperity index vectors obtained by interpolating monthly electricity sales data for each industry. The power grid service interference characteristics include the number of fault repair work orders. and average repair time .

3. The method for predicting electricity sales and correcting deviations in a power grid according to claim 1, characterized in that, The adaptive baseline decomposition includes the following steps: Define the business baseline at time t The components are trend component g(t), seasonal component s(t), and holiday component. The weighted sum; The trend component g(t) is defined by a piecewise linear growth model: in, Based on the growth rate, Basic intercept, Adjust the vector for the growth rate. The time-varying point indicator vector, The intercept smoothing vector is used to maintain the continuity of the function; The seasonal component s(t) is defined by a Fourier series: in, For the basic cycle, Let be the order of the Fourier series. These are the Fourier coefficients; The holiday component Defined as: in, The impact matrix of holidays, This is a vector of the impact coefficients of holidays. With a mean of 0 and a variance of It follows a normal distribution.

4. The method for predicting electricity sales and correcting deviations in a power grid according to claim 3, characterized in that, The adaptive baseline decomposition also includes: Calculate the electricity sales sequence within the time window Local information entropy within ; Local information entropy Meteorological feature vector at the current moment By concatenating the features, we obtain the joint feature vector. ; Joint feature vectors Input the gating network and calculate the time-varying weight coefficients: in, and For learnable gating parameters, The dynamic weights correspond to the trend component, seasonal component, and holiday component, respectively.

5. The method for predicting electricity sales and correcting deviations in a power grid according to claim 4, characterized in that, The baseline prediction value The calculation formula is: 。 6. The method for predicting electricity sales and correcting deviations in a power grid according to claim 1, characterized in that, The graph structure edge weight The construction methods include: constructing static weights based on the direct consumption coefficients in the industry input-output table, and / or constructing dynamic weights based on the Pearson correlation coefficients of the historical electricity sales sequence of each node, and setting a threshold for the correlation coefficients to truncate them in order to maintain the sparsity of the graph structure. The feature propagation formula for the graph convolutional network is: in, To add a self-loop adjacency matrix, For degree matrix, For learnable weight matrix, It is a non-linear activation function. This represents the feature matrix of the nodes in the l-th layer of the graph convolutional network. = .

7. The method for predicting electricity sales and correcting deviations in a power grid according to claim 1, characterized in that, The open-loop prediction results The calculation formula is: 。 8. The method for predicting electricity sales and correcting deviations in a power grid according to claim 1, characterized in that, The multidimensional deviation state vector is calculated using the following formula: in This is the deviation from the previous day. This is a deviation from the same day last week. The mean deviation within the sliding window. The first difference is the deviation; The error hidden state vector is calculated as follows: The formula for calculating the dynamic correction Δt generated by the deep policy network is as follows: in, Here are the weight matrix and bias vector of the hidden layer. The weight matrix and bias vector of the output layer. To correct the amplitude scaling factor.

9. A power grid-based electricity sales forecasting and deviation closed-loop correction device, characterized in that, include: The feature construction module acquires refined meteorological features, industry-specific macroeconomic features, and power grid business interference features from the power grid operation side, and performs spatiotemporal alignment processing to construct a comprehensive feature vector at time t. ; Adaptive baseline decomposition module, based on historical electricity sales sequence Baseline predictions for time t are generated through adaptive baseline decomposition. ; The spatial feature extraction module constructs a graph structure with administrative power supply bureaus and key industries as nodes and economic dependencies as edges. V and E represent the sets of nodes and edges, respectively, and the feature vectors of each node are... Mapping to a graph structure, spatial correlation features are extracted using graph convolutional networks. ; The multi-source regression prediction module will integrate spatial correlation features. The historical feature sequence is input into the Transformer encoder, which uses a multi-head self-attention mechanism for nonlinear coupled feature extraction and outputs refined prediction values. ; The integrated prediction module trains a meta-learner based on validation set data to generate dynamic weight coefficients. For baseline predicted values Compared with detailed forecasts We perform weighted ensemble to obtain open-loop prediction results. ; The closed-loop feedback correction module constructs a multi-dimensional deviation state vector based on historical prediction deviation sequences. The error hidden state vector is obtained by encoding using a gated recurrent unit (GRU) network. The error hidden state vector is input into the deep policy network to generate dynamic correction values. The dynamic correction is added to the open-loop prediction result to obtain the final electricity sales forecast. .

10. A computer-readable storage medium, characterized in that, It stores computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, it can implement the steps of the method as described in any one of claims 1-8.

Images