A method for predicting net load considering distributed photovoltaic access on load side

Abstract

The application provides a kind of net load prediction method considering distributed photovoltaic access on load side, belong to power system load prediction field;Solve the problem that traditional direct prediction method is difficult to split the independent characteristics of power consumption and photovoltaic output after distributed photovoltaic access on load side, resulting in low net load prediction accuracy;The method comprises the following steps: data acquisition and load-distributed photovoltaic decoupling;Data preprocessing and feature engineering;Distributed photovoltaic output prediction, including distributed photovoltaic output trend component and oscillation component prediction, to obtain the total point prediction value of distributed photovoltaic output;Actual power load prediction, including actual power load trend component and oscillation component prediction, to obtain the total point prediction value of actual power load;Net load prediction;Regional collaborative prediction;Evaluation and explanation and output the final net load prediction result;The application is applied to net load prediction after distributed photovoltaic access on load side.

Description

Technical Field

[0001] This application relates to the field of power system load forecasting technology, and in particular to a net load forecasting method that takes into account load-side distributed photovoltaic access. Background Technology

[0002] With the advancement of the "dual carbon" target, the penetration rate of distributed photovoltaic (PV) power on the load side is rapidly increasing. However, its output is highly random, influenced by meteorological factors such as sunlight and temperature. Furthermore, user electricity load exhibits complex fluctuations due to differences in electricity consumption habits and equipment types. This results in the load-side net load (the difference between load and PV power generation) not having the same fluctuation patterns as traditional loads or PV output. Currently, the monitoring and prediction of load-side net load mostly rely on in-home electricity monitoring devices, which do not distinguish between distributed PV and user load, but only monitor the net load, directly modeling and predicting the net load as a whole. This approach fails to adequately separate the independent characteristics of electricity consumption and PV output. Since the fluctuation characteristics of load and distributed PV are completely different, directly modeling and predicting the net load as a whole ignores the coupling relationship between PV output and load, forcibly merging them into a single net load prediction method. This is equivalent to using a unified model to model two completely different types of data, limiting its prediction accuracy. Other methods propose monitoring distributed PV and load separately, but this requires additional monitoring equipment, increasing operating costs and making it difficult to cover dispersed users such as residential communities. Meanwhile, traditional net load point prediction models are trained based on fixed parameters, making it difficult to adapt to changes such as increased photovoltaic penetration and the access of new loads, and they lack fault tolerance for abnormal scenarios; moreover, most of them are day-ahead predictions, which are difficult to respond to short-term fluctuations in real time, and the output net load point prediction results are difficult to match the real-time operation requirements of the power grid.

[0003] Therefore, there is an urgent need for a net load forecasting technology that can achieve efficient decoupling of load and power data and high forecasting accuracy without additional hardware, so as to improve the reliability of load forecasting in load-side distributed photovoltaic access scenarios. Summary of the Invention

[0004] To address the issue of low net load forecasting accuracy caused by the difficulty of separating the independent characteristics of power consumption and photovoltaic output in traditional direct forecasting methods after load-side distributed photovoltaic (PV) integration, this application proposes a net load forecasting method that considers load-side distributed PV integration.

[0005] The technical solution adopted in this application is: a net load forecasting method considering load-side distributed photovoltaic (PV) grid connection, comprising the following steps:

[0006] Step 1: Data Acquisition and Load-Distributed PV Decoupling: Decoupling the actual electricity load at the PV grid connection point from the distributed PV output;

[0007] Step 2: Data Preprocessing and Feature Engineering: The actual electricity load and distributed photovoltaic output obtained from the decoupling in Step 1 are preprocessed, and the predicted feature sets of actual electricity load and distributed photovoltaic output are constructed respectively.

[0008] Step 3: Distributed photovoltaic power output prediction: This includes the prediction of the trend component and oscillation component of distributed photovoltaic power output, to obtain the total predicted value of distributed photovoltaic power output.

[0009] Step 4: Actual electricity load forecast: This includes forecasting the trend component and oscillation component of the actual electricity load to obtain the total predicted value of the actual electricity load.

[0010] Step 5: Net load forecast: Combining the total output forecast of distributed photovoltaic power generation in Step 3 and the total actual electricity load forecast in Step 4, a preliminary net load forecast result is obtained, and the preliminary net load forecast result is checked and corrected for anomalies.

[0011] Step Six: Regional Collaborative Prediction: The data collected in Step One is assessed for completeness. For nodes with missing data, reference sources with matching characteristics are selected and weighted and fused to replace the net load prediction values ​​of the missing nodes, thereby achieving regional collaborative prediction.

[0012] Step 7: Evaluate and interpret, and output the final net load forecast results.

[0013] Furthermore, step one specifically includes:

[0014] Step 1.1: Data Acquisition: Based on the intelligent monitoring terminal of the user's grid connection point, collect the electricity load data and distributed photovoltaic output data of the photovoltaic grid connection point, as well as the surrounding meteorological data, and perform data filtering to retain a sample set with no less than 90% data integrity;

[0015] Step 1.2: Load-Distributed PV Decoupling: Based on the data selected in Step 1.1, an improved sequence-to-sequence model is used to separate the actual user electricity load from the distributed PV output. The improved sequence-to-sequence model is based on an encoder-decoder architecture. The encoder processes the multi-dimensional time-series information of the PV grid-connected points through multi-layer LSTM units, and outputs a hidden state sequence containing long-term time-series trends and short-term fluctuations. A time attention layer is embedded to assign high weights to key patterns and generate a context vector. The decoder combines this context vector with the predicted output of the previous time step, captures the time-series dependencies through LSTM units, and generates the decoupled actual electricity load and distributed PV output time-by-time through a fully connected layer.

[0016] Furthermore, step 1.2 specifically includes:

[0017] Step 1.2.1: Construction of multi-period load feature library: Divide the night into the late night period and the early morning period, calculate the load characteristics of the two periods respectively. The load characteristics include the mean and the standard deviation of fluctuation, and mark the low-amplitude stability characteristics of the late night period and the small-amplitude fluctuation characteristics of the early morning period respectively.

[0018] Step 1.2.2: Daytime decoupling modeling. For daytime periods, decoupling is achieved through "time period transition + attention mechanism", including:

[0019] 1) Time period identification: The daytime period is automatically divided into three core time periods: morning peak, noon peak, and evening peak;

[0020] 2) Load characteristic migration factor calculation: Define daytime period nighttime load mean migration factor and load fluctuation migration factor :

[0021] ;

[0022] ;

[0023] in, , These are the daytime periods of the same historical period. The load mean and load fluctuation standard deviation; , These represent the historical average nighttime load and the standard deviation of load fluctuation, respectively.

[0024] 3) Dynamic feature adaptation: through and Adjust the nighttime pure load sample to generate load baseline characteristics adapted to the current time period:

[0025] ;

[0026] in, As a load reference characteristic, For pure load samples, This is a random fluctuation term used to simulate the randomness of daytime loads;

[0027] 4) Power mutation attention mechanism: For net load mutation points, the irradiance change characteristics are enhanced through a time attention layer:

[0028] ;

[0029] in, Attention coefficient The total length of the entire time series. for Irradiance at any given time For the first The characteristic value at time;

[0030] Step 1.2.3: Use a dual verification mechanism to verify the decoupled daytime load and nighttime load, including: trend verification and residual verification;

[0031] Step 1.2.4: Decoupling verification.

[0032] Furthermore, step two specifically includes:

[0033] Step 2.1: Adaptive data decomposition and stabilization processing:

[0034] The actual electricity load is decomposed using a two-dimensional scenario decomposition strategy of "wavelet entropy-load fluctuation index": for high fluctuation scenarios, an improved variational mode decomposition is adopted to dynamically optimize the number of modes and the penalty factor to capture the characteristics of sudden load changes; for medium fluctuation scenarios, an improved local mean decomposition is adopted to decompose the load into a trend component and 3-5 product function components; for low fluctuation scenarios, a variational mode empirical wavelet transform is adopted to retain the trend component and 1-2 core oscillation components.

[0035] The distributed photovoltaic output is decomposed using enhanced empirical wavelet transform: the photovoltaic sequence is decomposed by adaptive Fourier spectrum segmentation technology, where the low frequency band corresponds to the trend component reflecting the annual / quarterly change of irradiance intensity, the mid frequency band corresponds to the sub-oscillation component reflecting the gradual change of intraday irradiance, and the high frequency band corresponds to the micro-oscillation component reflecting the instantaneous fluctuations such as rapid cloud cover.

[0036] Step 2.2: Data cleaning and standardization, including missing value imputation and data standardization processing;

[0037] Step 2.3: Feature construction, constructing prediction feature sets for actual electricity load and distributed photovoltaic output respectively:

[0038] Actual electricity load forecast feature set: including historical data, time characteristics, and temperature of the decomposed trend and oscillation components;

[0039] Distributed photovoltaic power output prediction feature set: includes historical data of the decomposed trend component and oscillation component, meteorological characteristics, and time characteristics.

[0040] Furthermore, step three specifically includes:

[0041] Step 3.1: Distributed photovoltaic power output trend component prediction:

[0042] The XGBoost-meteorological fusion model is used to predict the trend components of distributed photovoltaic power output. The XGBoost-meteorological fusion model includes a multi-source feature input layer, an XGBoost gradient boosting calculation layer, and a fusion output optimization layer. The input features of the XGBoost-meteorological fusion model include historical data of distributed photovoltaic power output trend components, core meteorological features, seasonal codes, and annual power output variation trends.

[0043] Step 3.2: Prediction of the oscillation component of distributed photovoltaic power output:

[0044] An attention-based LSTM model is adopted, which deeply integrates meteorological features and strengthens the driving role of meteorological features in photovoltaic power output prediction by dynamically allocating feature weights: the input layer inputs historical data of distributed photovoltaic power output oscillation components, real-time irradiance, irradiance change rate, cloud cover rate, and module temperature change rate; and sets up 3 layers of LSTM units, an irradiance attention layer, and a fully connected output layer.

[0045] Step 3.3: The sum of the predicted values ​​of the photovoltaic output trend component and the predicted values ​​of multiple distributed photovoltaic output oscillation components is used to obtain the total predicted value of distributed photovoltaic output.

[0046] Furthermore, step four specifically includes:

[0047] Step 4.1: Prediction of actual electricity load trend components:

[0048] An enhanced LSSVM model is used to predict the actual electricity load trend component. The input features of the enhanced LSSVM model include historical data of the actual electricity load trend component, average daily temperature, seasonal coding, user type features, and user behavior features.

[0049] Step 4.2: Prediction of the oscillation component of actual electricity load:

[0050] A spatiotemporal attention LSTM model is used to predict the oscillation component of actual electricity load. The spatiotemporal attention LSTM model includes an input layer, 2-3 layers of LSTM units, a spatiotemporal attention layer, and a fully connected output layer. The input layer takes into account historical data of the actual electricity load oscillation component, time period codes, and intraday peak and valley markings.

[0051] Step 4.3: The sum of the predicted values ​​of the actual electricity load trend component and the predicted values ​​of multiple actual electricity load oscillation components is used to obtain the total predicted value of the actual electricity load.

[0052] Furthermore, step five specifically includes:

[0053] Step 5.1: Net load point prediction calculation;

[0054] Step 5.2: Perform anomaly verification and correction on the preliminary net load forecast results, including:

[0055] 1) Amplitude verification, used to ensure that the absolute value of the net load is within a reasonable range;

[0056] 2) Mutation verification: The difference between adjacent time points must meet the set threshold. If the threshold is exceeded, the prediction results of the distributed photovoltaic power output oscillation component or the actual power load oscillation component need to be checked back. The attention weight of the model used to predict the distributed photovoltaic power output oscillation component or the actual power load oscillation component should be adjusted and then recalculated.

[0057] Step 5.3: Spatiotemporal correlation processing of net load forecast results.

[0058] Furthermore, step six specifically includes:

[0059] Step 6.1: Reference source filtering;

[0060] Step 6.2: Weight coefficient optimization:

[0061] Construct a bi-objective optimization function By adjusting the weight coefficients of each reference source Make the bi-objective optimization function Minimize the value:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] in, The target value for optimizing the weighting coefficients is used to evaluate the rationality of the weighting allocation. For continuously weighted coverage values, It is a continuous graded probability score. For the first The first reference source The weight of each indicator For the first The first reference source The standardized value of the indicator; For historical cycle number, For the first One reference source Forecast values ​​for the time period, For the first One reference source The corresponding actual value for the time period;

[0067] Each reference source is used first. calculate The initial value, then based on The relative proportions yield the initial weights. The initial weights are substituted in during the first iteration. The gradient descent algorithm is used to update the weight coefficients. The formula for the next iteration is:

[0068] ;

[0069] in, The learning rate; Indicates the first Each reference source passed through The weight coefficients are updated after each subgradient descent iteration, and the iteration is repeated until... No longer decreasing, meaning the total number of iterations has been reached. After this, the final updated weights are obtained. That is The optimal weighting coefficients for each reference source;

[0070] Step 6.3: Generate collaborative results: Generate the predicted values ​​of photovoltaic output and load points for missing nodes by weighted fusion of reference source intervals.

[0071] Furthermore, the criteria for selecting reference sources in step 6.1 are as follows:

[0072] Select 3-5 reference sources from photovoltaic grid-connected sites within the region with data integrity of no less than 95%, which must meet the following requirements:

[0073] Geographically, the straight-line distance between the reference source and the missing node does not exceed 5km and the elevation difference does not exceed 100m;

[0074] Regarding meteorological correlation, the correlation coefficients of irradiance and temperature between the reference source and the missing node are not less than 0.7;

[0075] In terms of load characteristics, the load type in the area where the reference source is located is consistent with that of the missing node, and the load peak-valley time deviation does not exceed 1 hour;

[0076] For time period matching, the net load characteristics of the reference source and the missing node in the same time period must be consistent, and the following conditions must be met: the mean deviation of net load in the same time period is ≤10%, and the standard deviation deviation of net load fluctuation in the same time period is ≤20%.

[0077] Furthermore, in step seven, the mean absolute error, root mean square error, and mean absolute percentage error are used to evaluate the effectiveness of the net load prediction results after anomaly verification and correction and spatiotemporal correlation processing, and the SHAP value of each feature is calculated based on the Shapley value principle of game theory.

[0078] The advantages of this application over the prior art are as follows:

[0079] 1. Low monitoring cost and flexible deployment: Only a single monitoring point needs to be deployed at the photovoltaic grid connection point. There is no need to install distributed photovoltaic or separate equipment metering devices, which reduces hardware costs by more than 60%. It is suitable for old communities, small industrial and commercial sites and other scenarios where it is difficult to deploy dedicated metering equipment on a large scale.

[0080] 2. Precise data decoupling: Through load-distributed photovoltaic decoupling and attention mechanism, it effectively solves the prediction bias problem caused by the traditional method of "aliasing modeling", and is more adaptable to scenarios such as sudden changes in photovoltaic output and sudden load increases;

[0081] 3. High prediction accuracy: Separate prediction models are built for load and photovoltaics, replacing the traditional unified modeling method. The point prediction error is significantly reduced, which can meet the needs of refined power grid dispatch.

[0082] 4. Wide adaptability to various scenarios: It supports multiple user types such as residential, commercial, and small industrial users, is compatible with various types of distributed photovoltaics, and solves the prediction problem of missing data nodes through regional collaborative prediction, making it more widely applicable. Attached Figure Description

[0083] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0084] Figure 1 This is an overall flowchart of the method provided in the embodiments of this application;

[0085] Figure 2 A comparison chart of non-intrusive load-photovoltaic decoupling results provided in the embodiments of this application;

[0086] Figure 3 This is a non-intrusive load-photovoltaic decoupling error analysis diagram provided in the embodiments of this application;

[0087] Figure 4 A flowchart for predicting distributed photovoltaic output incorporating meteorological attention is provided for embodiments of this application.

[0088] Figure 5 A comparison chart of predicted and actual net load values ​​provided in the embodiments of this application. Detailed Implementation

[0089] like Figures 1 to 5As shown, this application provides a net load forecasting method considering load-side distributed photovoltaic (PV) access. Relying solely on grid-connected data containing distributed PV loads, combined with meteorological data and load and PV feature mining, it constructs a multi-period pure load feature library. Utilizing a "period migration factor," it dynamically adapts to day-night load differences and embeds a power mutation attention mechanism to capture key fluctuations. This solves the technical challenge of decoupling dual variables from single-source data, eliminating the need for additional metering devices and achieving accurate decoupling and identification of actual user electricity load and distributed PV output. Furthermore, by performing point forecasts on both electricity load and PV output, it ultimately provides a net load point forecasting method that balances forecast accuracy, scenario adaptability, and feature interpretability, providing reliable technical support for grid dispatching and risk management in scenarios with high proportions of load-side distributed PV access.

[0090] like Figure 1 As shown, the net load forecasting method of this application includes the following steps:

[0091] Step 1: Data acquisition and load-distributed photovoltaic decoupling, including:

[0092] Step 1.1: Data Acquisition:

[0093] Based on the intelligent monitoring terminal at the user's grid connection point (which can be based on the existing power acquisition device), the active power of the grid connection point is collected and filtered. reactive power ,Voltage Current In addition to user type tags, distributed power supply type and installed capacity Meteorological data within a 5km radius (including irradiance) ,temperature Wind speed Cloud cover The 3σ criterion was used to identify outliers during data screening, and abnormal data was removed, retaining a sample set with at least 90% data integrity.

[0094] Step 1.2: Load-Distributed Photovoltaic Decoupling:

[0095] Based on the data selected above, the actual electricity load of users is separated using an improved sequence-to-sequence (Seq2Seq-TPA-LSTM) model. With distributed photovoltaic power output Satisfying the physical relationship: ,in This refers to the active power at the grid connection point, i.e., the net load.

[0096] The improved Sequence-to-Sequence (Seq2Seq-TPA-LSTM) model is based on an encoder-decoder architecture. Its core logic is as follows: The encoder processes multi-dimensional time-series information such as grid-connected power, meteorological data, and time period labels through multi-layer LSTM units, outputting a hidden state sequence containing long-term trends and short-term fluctuations. An embedded Time Attention (TPA) layer assigns high weights to key patterns such as morning peak hours and sudden photovoltaic changes, generating a context vector. The decoder combines this context vector with the predicted output from the previous moment, and uses LSTM units to capture temporal dependencies. A fully connected layer then generates decoupled actual electricity load and photovoltaic output time-by-time, achieving accurate separation of bivariate data from a single source. The specific implementation steps are as follows:

[0097] Step 1.2.1: Construction of a multi-period load characteristic library:

[0098] Extracting pure load samples by utilizing the characteristic of "no photovoltaic output at night". "No solar power output at night" refers to irradiance. hour, , The nighttime period is divided into late night (23:00-02:00) and early morning (02:00-05:00), and the load characteristics of the two periods are calculated separately, including the mean and standard deviation of fluctuation:

[0099] Mean: ,in This represents the average load during the late-night period. This represents the number of data points for the corresponding time period;

[0100] Standard deviation of fluctuation: ,in The standard deviation of load fluctuation during late-night hours;

[0101] Similarly, the average load during the early morning period can be calculated. and the standard deviation of load fluctuation during the early morning period .

[0102] Marked as "low-amplitude stable" late at night ( ) and the "slight fluctuations" in the early morning ( )characteristic.

[0103] Cross-scene supplement: Integrating rainy days (daytime) Data from time periods when photovoltaic output is negligible during the day, such as nighttime periods, are used to construct a load feature database covering both day and night, avoiding the limitations of a single nighttime sample.

[0104] Step 1.2.2: Daytime decoupling modeling:

[0105] For the daytime period (6:00-22:00), decoupling is achieved through "time period shifting + attention mechanism":

[0106] 1) Time Period Recognition: Automatically divides the day into three core time periods: morning peak (7:00-9:00), midday peak (12:00-14:00), and evening peak (18:00-22:00), and records them as follows: ,in Indicates daytime period. Indicates the morning rush hour. Indicates midday. This indicates the evening rush hour.

[0107] 2) Load characteristic migration factor calculation: Define daytime period nighttime load mean migration factor and load fluctuation migration factor :

[0108] ;

[0109] ;

[0110] in, , These are the daytime periods of the same historical period. The load mean and load fluctuation standard deviation; , These represent the historical average nighttime load and the standard deviation of load fluctuation, respectively.

[0111] For example, residential users typically... , Business lunchtime is usually .

[0112] This step adapts to the different load characteristics of different time periods by dividing the time period: electricity consumption varies significantly between different time periods (e.g., concentrated residential electricity consumption during the morning peak, high commercial electricity consumption plus photovoltaic output during midday, and overlapping residential and commercial loads during the evening peak). Combined with time period tags, different load mean migration factors can be applied accordingly. Load fluctuation migration factor .

[0113] 3) Dynamic feature adaptation: through and Adjust the nighttime pure load sample to generate load baseline characteristics adapted to the current time period:

[0114] ;

[0115] in, As a load reference characteristic, This is a random fluctuation term used to simulate the randomness of daytime loads.

[0116] 4) Power mutation attention mechanism: for net load mutation points ( , for Net load at all times (For installed capacity), the irradiance variation characteristics are enhanced through a time-attention layer:

[0117] ;

[0118] in, Attention coefficient (automatically adjusted to 0.02 in extreme weather conditions). The total length of the entire time series (i.e., the total number of moments contained in the time series data to be analyzed) enables the model to prioritize capturing sudden changes in photovoltaic output caused by cloud cover and other factors. for Irradiance at any given time For the first The characteristic value at time step.

[0119] Step 1.2.3: Employ a dual-verification mechanism to verify the decoupled daytime and nighttime loads, including:

[0120] 1) Trend verification: The daytime decoupled load must conform to the daytime trend of the user type (e.g., the peak of the morning rush hour for residents is 7:30-8:30), and the deviation from the trend of the same type of day and the same temperature in the same period of history should be ≤20%;

[0121] 2) Residual verification: To ensure the physical rationality of the decoupling results, the average daily load is defined. Compared with nighttime average load ratio It must fall within a reasonable range specific to the user type (residential user) Business users Industrial users ).

[0122] Step 1.2.4: Decoupling Verification:

[0123] The photovoltaic output obtained by decoupling must meet the following requirements. Daily average load after decomposition Compared with nighttime average load Deviation must meet If the deviation exceeds the limit, the model attention weights will be backtracked and adjusted to be re-decomposed.

[0124] Step 2: Data preprocessing and feature engineering, decoupled from the data obtained in Step 1. , Preprocessing is performed separately, including:

[0125] Step 2.1: Adaptive data decomposition and stabilization processing:

[0126] For the actual power load of users Decomposition: A two-dimensional scenario decomposition strategy of "wavelet entropy-load fluctuation index (LFI)" is adopted. For high-fluctuation scenarios, improved variational mode decomposition (IVMD) is used to dynamically optimize the number of modes and the penalty factor to capture the characteristics of sudden load changes; for medium-fluctuation scenarios, improved local mean decomposition (LMD) is used to decompose the load into a trend component and 3-5 product function components; for low-fluctuation scenarios, variational mode empirical wavelet transform (VMEWT) is used, which improves the decomposition efficiency by more than 25% and retains the trend component and 1-2 core oscillation components.

[0127] Output of distributed photovoltaic power Decomposition: Enhanced Empirical Wavelet Transform (EEWT) is used to decompose the photovoltaic sequence through adaptive Fourier spectrum segmentation technology. The low-frequency band corresponds to the trend component reflecting the annual / quarterly change in irradiance intensity, the mid-frequency band corresponds to the sub-oscillation component reflecting the gradual change in intraday irradiance, and the high-frequency band corresponds to the micro-oscillation component reflecting instantaneous fluctuations such as rapid cloud cover.

[0128] Step 2.2: Data cleaning and standardization, including:

[0129] Missing value imputation: A scenario-based strategy is adopted. When missing, fill in the gaps using the similarity interpolation method; When missing, fill in the missing information using a mapping method.

[0130] Data standardization: Using the Z-score formula:

[0131] ;

[0132] in, The mean, The standard deviation is used to eliminate differences in dimensions and ensure that feature weights are balanced when modeling two types of data.

[0133] Step 2.3: Feature construction, namely... and Constructing the predictive feature set:

[0134] Predictive feature set: includes historical data of decomposed trend and oscillation components, time features (year / quarter / month / day / hour encoding, day type, electricity peak and valley marking), and temperature;

[0135] Predictive feature set: includes historical data of decomposed trend and oscillation components, meteorological features (irradiance, temperature, cloud cover), and temporal features (seasonal coding, sunrise and sunset times).

[0136] Step 3: Distributed Photovoltaic Power Output Prediction: Based on the results obtained in step two The components are modeled and then superimposed to obtain the predicted point value. ,include:

[0137] Step 3.1: Trend component prediction:

[0138] Using the XGBoost-meteorological fusion model to... The trend component is used for prediction. The core structure of the XGBoost-meteorological fusion model includes a multi-source feature input layer, an XGBoost gradient boosting calculation layer, and a fusion output optimization layer. It strengthens the driving role of meteorological features in photovoltaic prediction through feature-level fusion. The input features of this model include... Historical data of trend components, core meteorological characteristics, seasonal codes, and annual power output variation trends;

[0139] XGBoost's parameter optimization is achieved through grid search: tree depth 3-10, learning rate 0.01-0.1, number of estimators 100-500, minimizing MAE, and after normalization, MAE is less than 0.06.

[0140] Step 3.2: Oscillation component prediction:

[0141] An attention-based LSTM model is employed, deeply integrating meteorological features. The driving role of meteorological features in photovoltaic power output prediction is strengthened through dynamic allocation of feature weights. The input layer includes historical data of oscillation components, real-time irradiance, irradiance change rate, cloud cover rate, and module temperature change rate. Three layers of LSTM units (128-256 nodes, Dropout = 0.3-0.6) and an irradiance attention layer (irradiance change rate > 50 W / (m²)) are configured. 2 •h) Increase feature weights and fully connected output layer.

[0142] Training strategy: Adam optimizer, learning rate 0.0001-0.001, combined with early stopping mechanism, normalized MAE is less than 0.04.

[0143] Step 3.3: Total points forecast:

[0144] Superposition formula: ;

[0145] in, This is the predicted total output value of distributed photovoltaic power generation. This represents the predicted value of the photovoltaic output trend component. The number of core oscillation components, For the first If the predicted value of each oscillation component is less than 0.09 after normalization of the total MAE, the meteorological feature input should be adjusted if this condition is not met.

[0146] Step 4: Actual power load of the user Prediction: Based on the results obtained in step two The components are modeled separately and then superimposed to obtain the predicted point values, including:

[0147] Step 4.1: Trend component prediction:

[0148] An enhanced LSSVM model is adopted, transforming the inequality constraints of traditional Support Vector Machine (SVM) into equality constraints. Model optimization improves the fitting accuracy of load trend components. Its input features include historical trend component data, daily average temperature, seasonal codes, user type features, and user behavior features. User behavior features are used to optimize trend component prediction, as they characterize the actual electricity consumption patterns and habits of power users, thus improving the accuracy of load forecasting. Model optimization employs Bayesian optimization, optimizing the radial basis function kernel parameters and regularization parameters. A distributed power source penetration rate feedback mechanism is introduced; the weight of the regularization parameter increases for every 5% increase in penetration rate. The accuracy requirement is a normalized MAE < 0.07.

[0149] Step 4.2: Oscillation component prediction:

[0150] A spatiotemporal attention LSTM model is adopted: the input layer inputs historical data of oscillation components, time period encoding, and intraday electricity peak and valley markings; 2-3 layers of LSTM units are set, with 64-128 nodes in each layer, and Dropout is introduced to prevent overfitting; spatiotemporal attention layer; fully connected output layer.

[0151] The training uses the Adam optimizer with a learning rate of 0.001-0.01. Combined with an early stopping mechanism, the normalized MAE does not exceed 0.05.

[0152] Step 4.3: Total points forecast:

[0153] Superposition formula: ;

[0154] in, This is the predicted total actual electricity load for the user. This represents the predicted value of the electricity consumption trend component. The number of core oscillation components, For the first If the predicted value of each oscillation component is less than 0.11 after normalization of the total MAE, and this condition is not met, the decomposition parameters are adjusted retrospectively.

[0155] Step 5: Net load forecasting, including:

[0156] Step 5.1: Net Load Point Prediction Calculation:

[0157] Net load point forecast Defined as the power that the grid side actually needs to supply or absorb, the calculation formula is:

[0158] ;

[0159] like This indicates that the user's actual electricity load exceeds the output of distributed photovoltaic power, and the power grid needs to supply electricity to the user.

[0160] like This indicates that the output of distributed photovoltaic power exceeds the actual electricity load of users, and the excess electricity is fed into the grid or stored in energy storage. It is marked as a "net output scenario" and is included separately in the reference for energy storage scheduling and grid absorption.

[0161] Step 5.2: Anomaly Validation and Correction of Prediction Results:

[0162] To ensure the physical reasonableness of the net load forecast results, new anomaly verification rules have been added, including:

[0163] 1) Amplitude verification, used to ensure that the absolute value of the net load is within a reasonable range:

[0164] In power supply scenarios, to avoid exceeding the user's maximum power demand, the following should be met: ,in This is the actual net load value. This represents the user's historical maximum load. If... Truncation correction is used: Record the excess amount at the same time. As a feedback feature for subsequent load model optimization, it is used to adjust the weights of user behavior features in the enhanced LSSVM model for load trend component prediction in step 4.1.

[0165] In the net output scenario, to avoid exceeding the maximum output of the photovoltaic installation, the following should be met: If the net output scenario Truncation correction is used: The sources of photovoltaic prediction bias are then identified and used to adjust the irradiance-related weights (including real-time irradiance and irradiance change rate) of the attention mechanism LSTM model for predicting photovoltaic oscillation components in step 3.2. When the source of bias is irradiance prediction error, the weight of the "irradiance change rate" feature in the model is increased to enhance the sensitivity to sudden changes in irradiance and reduce the resulting photovoltaic output deviation.

[0166] 2) Sudden change check: The difference between adjacent time points must meet the following requirements. If the threshold is exceeded, it is necessary to backtrack and check the prediction results of the distributed photovoltaic power output oscillation component or the actual electricity load oscillation component. After adjusting the attention weights of the model used to predict the distributed photovoltaic power output oscillation component or the actual electricity load oscillation component, the calculation should be performed: split the mutation contribution and calculate the photovoltaic mutation. and load mutation .like (PV-dominated), then increase the attention weight of the irradiance change rate in the model used to predict the PV oscillation component, and recalculate the net load after prediction; if If load is dominant, then the user behavior feature weights of the load trend component prediction model are increased, and the net load is recalculated after prediction. If it still exceeds the limit, a moving average smoothing is used to ensure that the fluctuations conform to physical constraints.

[0167] Step 5.3: Spatiotemporal correlation processing of net load forecast results, including:

[0168] Spatial correlation: Calculate the net load correlation coefficient between the photovoltaic grid-connected point to be predicted and adjacent photovoltaic grid-connected points (distance ≤ 5km) in the region. If the correlation coefficient is ≥ 0.6 (like adjacent users in the same community), the net load prediction value of the adjacent photovoltaic grid-connected points is used as an auxiliary feature for collaborative prediction. The net load prediction value of the adjacent photovoltaic grid-connected points with a correlation coefficient ≥ 0.6 is used as the input feature and integrated with the historical load, photovoltaic output, meteorological data and other features of the target photovoltaic grid-connected point itself. Different weights are assigned through an attention mechanism to improve the generalization ability of the net load prediction model of the target photovoltaic grid-connected point. This net load prediction model is a model that integrates the prediction of actual electricity load component, the prediction of distributed photovoltaic output component, anomaly verification and correction, and spatiotemporal correlation collaboration, and finally outputs the net load prediction value of the target photovoltaic grid-connected point at future times.

[0169] Time correlation: Divide the forecast period into granularities and output the average net load for each period. Net load fluctuation standard deviation This provides data support for "time period matching" of reference sources in regional collaborative forecasting.

[0170] Step Six: Regional Collaborative Prediction: The completeness of the original collected data is assessed. Data completeness = (Number of valid data entries / Total number of data entries to be collected) × 100%, where valid data refers to normal data that is not missing after outlier removal using the 3σ criterion in Step 1.1. Regional collaborative prediction is required when the original collected data is determined to be missing.

[0171] For example, if the irradiance data of a certain photovoltaic grid-connected point is missing by 10% (completeness <95%), the photovoltaic output cannot be accurately predicted due to the lack of key meteorological input. Therefore, it is necessary to select reference nodes with complete irradiance data within 5km through regional collaborative prediction, and use their irradiance data to replace the missing part to assist in decoupling and prediction.

[0172] This step specifically includes:

[0173] Step 6.1: Reference source filtering:

[0174] From photovoltaic grid-connected points with data integrity of no less than 95% within the region, select 3-5 reference sources that meet the following requirements: Geographically, the straight-line distance between the reference source and the missing node should not exceed 5km and the altitude difference should not exceed 100m; In terms of meteorological correlation, the correlation coefficients of irradiance and temperature between the reference source and the missing node should not be less than 0.7; In terms of load characteristics, the load type of the area where the reference source is located should be consistent with that of the missing node, and the peak-valley time deviation should not exceed 1 hour; In terms of time period matching, the net load characteristics of the reference source and the missing node should be consistent during the same time period, and should meet the following requirements: the mean deviation of net load during the same time period should be ≤10%, and the standard deviation deviation of net load fluctuation during the same time period should be ≤20%.

[0175] Step 6.2: Weight coefficient optimization:

[0176] Construct a bi-objective optimization function By adjusting the weight coefficients of each reference source Make the bi-objective optimization function Minimize the value of (i.e., optimize the weight allocation most effectively), where the bi-objective optimization function is... include:

[0177] CWC (Continuously Weighted Coverage): Used to reflect the comprehensive matching degree between the reference source and the missing site on multiple indicators, including geographical distance, altitude difference, meteorological correlation and load characteristics; the higher the matching degree, the larger the CWC value.

[0178] CRPS (Continuous Graded Probability Score): Calculated by comparing the predicted values ​​of the reference source with the actual observed values. The smaller the prediction bias, the smaller the CRPS value.

[0179] ;

[0180] ;

[0181] in, For the first The first reference source The weight of each indicator (initial value) ), For the first The first reference source The standardized value of the indicator; Historical cycle number (e.g., historical) (same time period) For the first One reference source Forecast values ​​for the time period, For the first One reference source The corresponding actual value for the time period.

[0182] Each reference source is used first. calculate The initial value, then based on The relative proportions yield the initial weights. ( The smaller the value, the better the reference source, and the larger the initial weight.

[0183] ;

[0184] ;

[0185] in, The target value for the weighting coefficients is used to evaluate the rationality of the weight allocation; the smaller the value, the better the weight allocation. Weighting coefficients of 0.6 and 0.4 are used to balance the influence of CWC and CRPS in the objective function.

[0186] First iteration, substitute initial weights The gradient descent algorithm is used to update the weight coefficients. The formula for the next iteration is:

[0187] ;

[0188] in, The learning rate (step size); Indicates the first Each reference source passed through The weight coefficients are updated after each subgradient descent iteration, and the iteration is repeated until... No longer decreasing, meaning the total number of iterations has been reached. After this, the final updated weights are obtained. That is The optimal weighting coefficients for each reference source.

[0189] Step 6.3: Generate collaborative results:

[0190] By weighted fusion of reference source intervals, the predicted values ​​of photovoltaic output and load points for missing nodes are generated:

[0191] ;

[0192] in, The predicted value (photovoltaic output or load) for the missing node. For the first The weighting coefficients of each reference source, For the first The predicted values ​​from each reference source.

[0193] Step Seven: Assessment and Interpretation, including:

[0194] Step 7.1: Validity Assessment:

[0195] The forecasting performance is evaluated using three core metrics: mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE).

[0196] If MAE > 0.1, RMSE > 0.15, or MAPE > 10%, backtrack to optimize model parameters (such as adjusting the number of nodes in LSTM and the learning rate of XGBoost).

[0197] Step 7.2: Feature Interpretability Analysis (SHAP Value):

[0198] The Shapley value (SHAP) of each feature is calculated based on the Shapley value principle of game theory: For each sample, first calculate the prediction difference when each feature is "present" and "absent", then iterate through all feature combinations and take the average marginal contribution as the SHAP value of that sample; next, take the absolute value of the feature SHAP value of the entire sample, calculate its proportion of the total SHAP value, and statistically analyze the distribution range of this proportion (e.g., the 25%-75% quantile range) to obtain the contribution interval, and sort them according to feature importance:

[0199] Photovoltaic output forecast: Irradiance contribution 35%-45%, temperature contribution 20%-30%, time characteristic contribution 15%-25%, historical output contribution 10%-15%;

[0200] Load forecasting: temperature contributes 25%-35%, time characteristics contribute 30%-40%, user type characteristics contribute 15%-25%, and historical load contributes 10%-15%;

[0201] Analysis of the influence direction of characteristics: when the irradiance exceeds 800 W / m 2 When the temperature is within 35℃, the prediction uncertainty decreases; when the temperature exceeds 35℃, the prediction uncertainty increases. Based on this result, data acquisition and feature weights are optimized.

[0202] The contribution of the above features is ranked according to their importance. The main features are retained. For example, cloud cover is not included in the contribution analysis because it is of low importance in photovoltaic power output forecasting. User behavior features are not included in the contribution analysis because they are of lower importance than other features in load forecasting.

[0203] This application proposes a net load forecasting method that decouples distributed photovoltaic (PV) output and load by considering the "time migration factor." It is applicable to scenarios such as industrial parks, urban distribution networks, and residential communities where a high proportion of load-side distributed PV (such as user rooftop PV and industrial and commercial distributed PV) are connected. It can achieve accurate forecasting of distributed power output, actual user electricity load, and grid-side net load without the need to install additional metering equipment on the user side or power source side, providing quantitative basis for grid dispatching, energy storage optimization, and risk management.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A net load forecasting method considering load-side distributed photovoltaic (PV) grid connection, characterized in that: Includes the following steps: Step 1: Data Acquisition and Load-Distributed PV Decoupling: Decoupling the actual electricity load at the PV grid connection point from the distributed PV output; Step 2: Data Preprocessing and Feature Engineering: The actual electricity load and distributed photovoltaic output obtained from the decoupling in Step 1 are preprocessed, and the predicted feature sets of actual electricity load and distributed photovoltaic output are constructed respectively. Step 3: Distributed photovoltaic power output prediction: This includes the prediction of the trend component and oscillation component of distributed photovoltaic power output, to obtain the total predicted value of distributed photovoltaic power output. Step 4: Actual electricity load forecast: This includes forecasting the trend component and oscillation component of the actual electricity load to obtain the total predicted value of the actual electricity load. Step 5: Net load forecast: Combining the total output forecast of distributed photovoltaic power generation in Step 3 and the total actual electricity load forecast in Step 4, a preliminary net load forecast result is obtained, and the preliminary net load forecast result is checked and corrected for anomalies. Step Six: Regional Collaborative Prediction: The data collected in Step One is assessed for completeness. For nodes with missing data, reference sources with matching characteristics are selected and weighted and fused to replace the net load prediction values ​​of the missing nodes, thereby achieving regional collaborative prediction. Step 7: Evaluate and interpret, and output the final net load forecast results.

2. The net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 1, characterized in that: Step one specifically includes: Step 1.1: Data Acquisition: Based on the intelligent monitoring terminal of the user's grid connection point, collect the electricity load data and distributed photovoltaic output data of the photovoltaic grid connection point, as well as the surrounding meteorological data, and perform data filtering to retain a sample set with no less than 90% data integrity; Step 1.2: Load-Distributed PV Decoupling: Based on the data selected in Step 1.1, an improved sequence-to-sequence model is used to separate the actual user electricity load from the distributed PV output. The improved sequence-to-sequence model is based on an encoder-decoder architecture. The encoder processes the multi-dimensional time-series information of the PV grid-connected points through multi-layer LSTM units, and outputs a hidden state sequence containing long-term time-series trends and short-term fluctuations. A time attention layer is embedded to assign high weights to key patterns and generate a context vector. The decoder combines this context vector with the predicted output of the previous time step, captures the time-series dependencies through LSTM units, and generates the decoupled actual electricity load and distributed PV output time-by-time through a fully connected layer.

3. The net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 2, characterized in that: Step 1.2 specifically includes: Step 1.2.1: Construction of multi-period load feature library: Divide the night into the late night period and the early morning period, calculate the load characteristics of the two periods respectively. The load characteristics include the mean and the standard deviation of fluctuation, and mark the low-amplitude stability characteristics of the late night period and the small-amplitude fluctuation characteristics of the early morning period respectively. Step 1.2.2: Daytime decoupling modeling. For daytime periods, decoupling is achieved through "time period transition + attention mechanism", including: 1) Time period identification: The daytime period is automatically divided into three core time periods: morning peak, noon peak, and evening peak; 2) Load characteristic migration factor calculation: Define daytime period nighttime load mean migration factor and load fluctuation migration factor : ； ； in, , These are the daytime periods of the same historical period. The load mean and load fluctuation standard deviation; , These represent the historical average nighttime load and the standard deviation of load fluctuation, respectively. 3) Dynamic feature adaptation: through and Adjust the nighttime pure load sample to generate load baseline characteristics adapted to the current time period: ； in, As a load reference characteristic, For pure load samples, This is a random fluctuation term used to simulate the randomness of daytime loads; 4) Power mutation attention mechanism: For net load mutation points, the irradiance change characteristics are enhanced through a time attention layer: ； in, Attention coefficient The total length of the entire time series. for Irradiance at any given time For the first Irradiance at any given time; Step 1.2.3: Use a dual verification mechanism to verify the decoupled daytime load and nighttime load, including: trend verification and residual verification; Step 1.2.4: Decoupling verification.

4. The net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 3, characterized in that: Step two specifically includes: Step 2.1: Adaptive data decomposition and stabilization processing: The actual electricity load is decomposed using a two-dimensional scenario decomposition strategy of "wavelet entropy-load fluctuation index": for high fluctuation scenarios, an improved variational mode decomposition is used to dynamically optimize the number of modes and the penalty factor to capture the characteristics of sudden load changes; for medium fluctuation scenarios, an improved local mean decomposition is used to decompose the load into a trend component and 3-5 product function components; for low fluctuation scenarios, a variational mode empirical wavelet transform is used to retain the trend component and 1-2 core oscillation components. The distributed photovoltaic output is decomposed using enhanced empirical wavelet transform: the photovoltaic sequence is decomposed by adaptive Fourier spectrum segmentation technology, where the low frequency band corresponds to the trend component reflecting the annual / quarterly change of irradiance intensity, the mid frequency band corresponds to the sub-oscillation component reflecting the gradual change of intraday irradiance, and the high frequency band corresponds to the micro-oscillation component reflecting the instantaneous fluctuation of cloud shading. Step 2.2: Data cleaning and standardization, including missing value imputation and data standardization processing; Step 2.3: Feature construction, constructing prediction feature sets for actual electricity load and distributed photovoltaic output respectively: Actual electricity load forecast feature set: including historical data, time characteristics, and temperature of the decomposed trend and oscillation components; Distributed photovoltaic power output prediction feature set: includes historical data of the decomposed trend component and oscillation component, meteorological characteristics, and time characteristics.

5. The net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 4, characterized in that: Step three specifically includes: Step 3.1: Component Prediction of Distributed Photovoltaic Output Trend: The XGBoost-meteorological fusion model is used to predict the trend components of distributed photovoltaic power output. The XGBoost-meteorological fusion model includes a multi-source feature input layer, an XGBoost gradient boosting calculation layer, and a fusion output optimization layer. The input features of the XGBoost-meteorological fusion model include historical data of distributed photovoltaic power output trend components, core meteorological features, seasonal codes, and annual power output variation trends. Step 3.2: Prediction of the oscillation component of distributed photovoltaic power output: An attention-based LSTM model is adopted, which deeply integrates meteorological features and strengthens the driving role of meteorological features in photovoltaic power output prediction by dynamically allocating feature weights: the input layer inputs historical data of distributed photovoltaic power output oscillation components, real-time irradiance, irradiance change rate, cloud cover rate, and module temperature change rate; and sets up 3 layers of LSTM units, an irradiance attention layer, and a fully connected output layer. Step 3.3: The sum of the predicted values ​​of the photovoltaic output trend component and the predicted values ​​of multiple distributed photovoltaic output oscillation components is used to obtain the total predicted value of distributed photovoltaic output.

6. The net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 5, characterized in that: Step four specifically includes: Step 4.1: Prediction of actual electricity load trend components: An enhanced LSSVM model is used to predict the actual electricity load trend component. The input features of the enhanced LSSVM model include historical data of the actual electricity load trend component, average daily temperature, seasonal coding, user type features, and user behavior features. Step 4.2: Prediction of the oscillation component of actual electricity load: A spatiotemporal attention LSTM model is used to predict the oscillation component of actual electricity load. The spatiotemporal attention LSTM model includes an input layer, 2-3 layers of LSTM units, a spatiotemporal attention layer, and a fully connected output layer. The input layer takes into account historical data of the actual electricity load oscillation component, time period codes, and intraday peak and valley markings. Step 4.3: The sum of the predicted values ​​of the actual electricity load trend component and the predicted values ​​of multiple actual electricity load oscillation components is used to obtain the total predicted value of the actual electricity load.

7. A net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 6, characterized in that: Step five specifically includes: Step 5.1: Net load point prediction calculation; Step 5.2: Perform anomaly verification and correction on the preliminary net load forecast results, including: 1) Amplitude verification, used to ensure that the absolute value of the net load is within a reasonable range; 2) Mutation verification: The difference between adjacent time points must meet the set threshold. If the threshold is exceeded, the prediction results of the distributed photovoltaic power output oscillation component or the actual power load oscillation component need to be checked back. The attention weight of the model used to predict the distributed photovoltaic power output oscillation component or the actual power load oscillation component should be adjusted and then recalculated. Step 5.3: Spatiotemporal correlation processing of net load forecast results.

8. The net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 7, characterized in that: Step six specifically includes: Step 6.1: Reference source filtering; Step 6.2: Weight coefficient optimization: Construct a bi-objective optimization function By adjusting the weight coefficients of each reference source Make the bi-objective optimization function Minimize the value: ； ； ； ； in, The target value for optimizing the weighting coefficients is used to evaluate the rationality of the weighting allocation. For continuously weighted coverage values, It is a continuous graded probability score. For the first The first reference source The weight of each indicator For the first The first reference source The standardized value of the indicator; For historical cycle number, For the first One reference source Forecast values ​​for the time period, For the first One reference source The corresponding actual value for the time period; Each reference source is used first. calculate The initial value, then based on The relative proportions yield the initial weights. The initial weights are substituted in during the first iteration. The gradient descent algorithm is used to update the weight coefficients. The formula for the next iteration is: ； in, The learning rate; Indicates the first Each reference source passed through The weight coefficients are updated after each subgradient descent iteration, and the iteration is repeated until... No longer decreasing, meaning the total number of iterations has been reached. After this, the final updated weights are obtained. That is The optimal weighting coefficients for each reference source; Step 6.3: Generate collaborative results: Generate the photovoltaic output and load point prediction values ​​for missing nodes by weighted fusion of reference source intervals.

9. A net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to claim 8, characterized in that: The criteria for selecting reference sources in step 6.1 are as follows: Select 3-5 reference sources from photovoltaic grid-connected sites within the region with data integrity of no less than 95%, which must meet the following requirements: Geographically, the straight-line distance between the reference source and the missing node does not exceed 5km and the elevation difference does not exceed 100m; Regarding meteorological correlation, the correlation coefficients of irradiance and temperature between the reference source and the missing node are not less than 0.7; In terms of load characteristics, the load type in the area where the reference source is located is consistent with that of the missing node, and the load peak-valley time deviation does not exceed 1 hour; For time period matching, the net load characteristics of the reference source and the missing node in the same time period must be consistent, and the following conditions must be met: the mean deviation of net load in the same time period is ≤10%, and the standard deviation deviation of net load fluctuation in the same time period is ≤20%.

10. A net load forecasting method considering load-side distributed photovoltaic (PV) grid connection according to any one of claims 1-9, characterized in that: In step seven, the mean absolute error, root mean square error, and mean absolute percentage error are used to evaluate the effectiveness of the net load prediction results after anomaly verification and correction and spatiotemporal correlation processing, and the SHAP value of each feature is calculated based on the Shapley value principle of game theory.

Images