Low-voltage distributed photovoltaic output and charging station power combined intelligent prediction method

Abstract

The invention discloses a low-voltage distributed photovoltaic output and charging station power combined intelligent prediction method, relates to the technical field of power prediction, and effectively solves the coupling problem of photovoltaic and charging load prediction in a low-voltage power distribution network through collaborative innovation of a multi-dimensional technology. The characteristic engineering based on the weather-load joint sensitivity coefficient quantifies the dynamic influence of the weather sudden change on the two sides of the source load, and solves the prediction deviation caused by the characteristic space splitting in the traditional method. Secondly, the space-time attention guided joint prediction model extracts local meteorological features through a convolutional network to model power grid topological association; and finally, a closed-loop feedback system of error monitoring, root cause analysis, incremental learning and effect verification is constructed, prediction system self-healing is realized through online parameter fine tuning and model hot updating, the three technologies form an organic whole, prediction precision and real-time performance are significantly improved, and reliable decision support is provided for a high-proportion new energy access scene.

Description

Technical Field

[0001] The present invention relates to the technical field of power prediction, and in particular to a method for intelligently predicting low-voltage distributed photovoltaic output and charging station power. Background Art

[0003] Current research mostly adopts an isolated modeling paradigm: photovoltaic forecasting focuses on the time series extrapolation of meteorological factors, but ignores the power reaction of charging behavior on the local power grid; charging load forecasting relies on historical statistical laws, but does not quantify the transmission effect of weather mutations on the user's travel chain. More importantly, the existing methods have three limitations: first, the data level lacks multimodal fusion of meteorology-user behavior-grid topology, and cannot characterize the conflict scenario of "sharp drop in photovoltaic output" and "surge in charging demand" in cloudy weather; second, the model level is limited by the static weight distribution mechanism, and it is difficult to capture the dynamic reconstruction characteristics of the spatiotemporal correlation between source and load in extreme events such as typhoons; third, the correction level generally adopts an offline calibration strategy, which lags behind the rapid evolution of the real-time operating status of the distribution network. The above defects have caused the prediction error of existing technologies in actual projects to generally exceed 15%, seriously restricting the reliability of grid dispatching decisions. Summary of the invention

[0004] The purpose of the present invention is to provide a method for joint intelligent prediction of low-voltage distributed photovoltaic output and charging station power, propose a multi-source feature coupling framework driven by meteorological sensitivity, incorporate hidden variables into a unified feature space for the first time, and overcome the problem of semantic alignment of heterogeneous data; secondly, design a joint prediction architecture guided by spatiotemporal attention to realize cross-modal correlation mining of photovoltaic output fluctuations and charging load peaks and valleys; finally, develop an online correction mechanism based on edge computing, and use the dynamic tracing technology of the error propagation path to enable the model to have correction capabilities in emergency scenarios to solve the problems raised in the prior art.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a method for jointly intelligently predicting low-voltage distributed photovoltaic output and charging station power, which constructs an information feature set between distributed photovoltaic output and user charging willingness under different environmental factors;

[0006] A joint intelligent prediction model of distributed photovoltaic output and charging station power is established based on information feature sets and electricity consumption feature data;

[0007] Establish performance correction mechanism for intelligent prediction models;

[0008] Combined prediction of distributed photovoltaic output and charging station power is carried out based on the intelligent prediction model of the subsidiary performance correction mechanism;

[0009] According to the above technical solution, the construction of the information feature set includes the following steps:

[0010] S101. Perform multi-source data collection and preprocessing;

[0011] Furthermore, the multi-source data includes meteorological data, photovoltaic data, and charging behavior data;

[0012] Furthermore, the preprocessing includes time alignment, outlier handling, and standardization;

[0013] S102. Calculate features for the data processed in S101 in combination with the sensitivity coefficient;

[0014] S103. Establish a feature screening mechanism to screen the features and obtain an information feature set.

[0015] According to the above technical solution, in S102, the dynamic impacts of meteorological variables on photovoltaic output and charging behavior are synchronously quantified through a single coefficient;

[0016] Furthermore, it includes two core terms: photovoltaic sensitivity term and charging elasticity term;

[0017] In S103, the screening of features refers to calculating the maximum information coefficient of each feature with the target variable, retaining the relevant features whose maximum information coefficient exceeds the set threshold, and manually reviewing the rationality of the features.

[0018] According to the above technical solution, the establishment of the intelligent prediction model includes the following steps:

[0019] S201. Establish an input module for the joint intelligent prediction model of distributed photovoltaic output and charging station power;

[0020] S202. Establish a feature extraction layer for the joint intelligent prediction model of distributed photovoltaic output and charging station power;

[0021] S203. Establish a spatio-temporal correlation model between distributed photovoltaic output and charging station power;

[0022] S204. Design a multi-objective loss function for the spatio-temporal correlation model;

[0023] S205. Form a dynamic learning mechanism.

[0024] According to the above technical solution, in step S201, the establishment of the input module includes data channel division and data standardization;

[0025] The data channel division includes a meteorological-photovoltaic channel and a user-grid channel;

[0026] Data standardization means implementing maximum-minimum normalization on the data of each channel to eliminate the dimension difference;

[0027] In S202, the establishment of the feature extraction layer refers to the extraction of local meteorological features;

[0028] It includes:

[0029] S202-1. Use a convolutional neural network to capture the local spatial correlation patterns of irradiance and cloud movement;

[0030] S202-2. Design a 3-layer convolutional structure, and the size of the convolutional kernel is adaptively adjusted along the time dimension;

[0031] S202-3. Adopt a bidirectional long short-term memory network to learn the intraday cycle characteristics of photovoltaic output and charging load, and introduce a weather type gating mechanism;

[0032] In S203, the spatio-temporal correlation modeling includes:

[0033] S203-1. Construct a power grid topology map: the nodes are distribution transformer substations, and the edges are line impedances;

[0034] S203-2. Calculate the power fluctuation propagation weights between nodes, and focus on capturing the cascading impact of photovoltaic output mutations on adjacent charging piles;

[0035] Furthermore, use a cross-modal attention mechanism to design a photovoltaic-charging load cross-attention module to quantify the mutual interference intensity between the two under extreme weather;

[0036] In S204, the design of the multi-objective loss function includes a main loss term and a physical constraint term;

[0037] Furthermore, the main loss term refers to the weighted sum of the mean square errors of photovoltaic and charging power predictions;

[0038] The physical constraint terms include power grid power balance constraints, temporal smoothing constraints, and mutual information maximization terms;

[0039] In S205, the dynamic learning mechanism includes a weather type adaptive learning rate and sliding window incremental training;

[0040] According to the above technical solutions, the establishment of the intelligent prediction model performance correction mechanism includes the following steps:

[0041] S301. Monitor the errors and diagnose anomalies of the intelligent prediction model;

[0042] S302. Online correct the intelligent prediction model according to the diagnosis results of S301.

[0043] According to the above technical solutions, in S301, it includes the following steps:

[0044] S301-1. Real-time calculation of multi-dimensional error metrics;

[0045] Among them, the multi-dimensional error index includes short-term error, long-term drift, and mutation detection;

[0046] S301-2. Conduct error root cause analysis;

[0047] It includes: feature contribution degree backtracking and associated device verification.

[0048] According to the above technical solution, in S302, the following steps are included:

[0049] S302-1. Adopt model hot update: Adopt the elastic weight consolidation algorithm to inject new samples while retaining historical knowledge to prevent catastrophic forgetting;

[0050] S302-2. Optimize the feature space;

[0051] It includes failure feature masking and derivative feature generation;

[0052] S302-3. Construct a feedback closed-loop;

[0053] It includes a short-term feedback closed-loop and a long-term feedback closed-loop.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] The present invention effectively solves the coupling problem of photovoltaic and charging load forecasting in low-voltage distribution networks through multi-dimensional technology collaborative innovation:

[0056] First, the feature engineering based on the meteorological-load joint sensitivity coefficient quantifies the dynamic impact of weather mutations on both the source and load sides, solving the prediction deviation caused by the fragmentation of the feature space in traditional methods;

[0057] Secondly, the joint prediction model guided by spatio-temporal attention extracts local meteorological features through a convolutional network to model the power grid topology association;

[0058] Finally, a closed-loop feedback system of "error monitoring - root cause analysis - incremental learning - effect verification" is constructed, and the prediction system is self-healed through online parameter fine-tuning and model hot update.

[0059] The three technologies form an organic whole, significantly improving the prediction accuracy and real-time performance, and providing reliable decision-making support for high-proportion new energy access scenarios. Brief Description of the Drawings

[0060] Figure 1 It is a comparison chart of photovoltaic output prediction for an embodiment of the present invention;

[0061] Figure 2 It is a charging power and correction effect diagram for an embodiment of the present invention. Detailed Embodiment

[0062] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0063] The present invention provides a technical solution for a joint intelligent prediction method of low-voltage distributed photovoltaic output and charging station power, including the following steps:

[0064] S1. Establish an information feature set between distributed photovoltaic output and user charging willingness under different environmental factors;

[0065] The construction of the information feature set includes the following steps:

[0066] S101. Perform multi-source data collection and preprocessing;

[0067] Furthermore, the multi-source data includes meteorological data, photovoltaic data, and charging behavior data;

[0068] Specifically:

[0069] The meteorological data includes real-time irradiance, unit: W / m 2 , cloud movement speed vector, decomposed into east-west and north-south direction components, unit: m / s, atmospheric transmittance, precipitation intensity, unit: mm / h;

[0070] The photovoltaic data includes the AC power output by the collected inverter, unit: kw, and the backplane temperature of the photovoltaic module, unit: °C;

[0071] The charging behavior data includes recording the start and stop times of the charging pile, the charging amount, unit: kwh, and the charging urgency index, calculated by the target power set by the user and the remaining charging time, unit: % / min;

[0072] Furthermore, the preprocessing includes time alignment, outlier processing, and normalization processing;

[0073] Specifically:

[0074] Time alignment: Use the dynamic time warping algorithm to align the timestamps of multi-source data to ensure that all variables are strictly synchronized at a 5-minute sampling interval, and the maximum allowable time deviation is ±30 seconds;

[0075] Outlier processing: Smoothly interpolate the abnormal data points where the irradiance mutation exceeds 300 W / m 2 to eliminate the influence of instantaneous equipment failures;

[0076] Standardization: Standardize the mean and variance of meteorological data to eliminate dimensional differences; for example, irradiance data is converted into standardized values centered on the historical mean and with standard deviation as the unit;

[0077] S102, perform characteristic calculation on the joint sensitivity coefficient of the data processed by S101; quantify the dynamic impact of meteorological variables on photovoltaic output and charging behavior through a single coefficient;

[0078] Furthermore, two core items are included: photovoltaic sensitivity item and charging elasticity item;

[0079] Specific:

[0080] Photovoltaic sensitive items: reflect the intensity of photovoltaic power fluctuation caused by unit meteorological changes, calculate the change of photovoltaic power at adjacent 5-minute time points, calculate the change of meteorological variables in the same period, such as the change of irradiance, and calculate the average ratio of the two with a sliding time window. The default 1-hour window contains 12 data points, which measures the immediate impact of meteorological mutations on photovoltaic power generation; specifically, it can be expressed as:

[0081]

[0082] Where GF is the photovoltaic sensitivity term; ΔP PV (t k ) is the time t k The difference in photovoltaic output at adjacent moments. Δx i (t k ) is the meteorological variable at time t k The change in the adjacent moments of the sliding time window. N is the number of sampling points in the sliding time window; ò is a very small constant used to avoid the denominator being zero;

[0083] Charging elasticity term: characterizes the regulatory effect of meteorological conditions on user charging demand; imposes a small disturbance near the historical mean of the meteorological variable, for example: the disturbance amplitude is 10% of the standard deviation of the variable; calculates the sensitivity of the charging urgency index to the meteorological variable by numerical differentiation method; specifically:

[0084]

[0085] Among them, CD is the charging elasticity term; U c (t) is the charging urgency function; μ i is the historical mean of meteorological variables; δ is the numerical differential perturbation step size;

[0086] S103, establishing a feature screening mechanism, screening features, and obtaining an information feature set;

[0087] In S103, the screening of features refers to calculating the maximum information coefficient between each feature and the target variable, and retaining the relevant features whose maximum information coefficient exceeds the set threshold. For example, the set threshold is 0.35, and the rationality of the features is manually reviewed.

[0088] S2. Establish a joint intelligent prediction model for distributed photovoltaic output and charging station power based on the information feature set and power consumption feature data;

[0089] Specifically, it includes the following steps:

[0090] S201. Establish an input module for the joint intelligent prediction model of distributed photovoltaic output and charging station power;

[0091] The establishment of the input module includes data channel division and data standardization;

[0092] Data channel division includes a meteorological-photovoltaic channel and a user-grid channel;

[0093] Specifically:

[0094] Meteorological-photovoltaic channel: Receive the constructed information feature set;

[0095] User-grid channel: Input the real-time power of the charging pile cluster, the historical sequence of grid node voltages, and the transformer load rate;

[0096] Data standardization refers to implementing maximum-minimum normalization on the data of each channel to eliminate the dimension difference;

[0097] S202. Establish a feature extraction layer for the joint intelligent prediction model of distributed photovoltaic output and charging station power;

[0098] The establishment of the feature extraction layer refers to performing local meteorological feature extraction;

[0099] It includes the following steps:

[0100] S202-1. Use a convolutional neural network to capture the local spatial correlation patterns of irradiance and cloud movement;

[0101] S202-2. Design a 3-layer convolutional structure, and the convolutional kernel size is adaptively adjusted along the time dimension. For example, it is 5×5 on sunny days and 3×3 on rainy days;

[0102] S202-3. Adopt a bidirectional long short-term memory network to learn the intraday cycle characteristics of photovoltaic output and charging load, and introduce a weather type gating mechanism, that is: adjust the forgetting gate weight according to the real-time weather state;

[0103] S203. Establish a spatio-temporal correlation model between distributed photovoltaic output and charging station power;

[0104] It includes the following steps:

[0105] S203-1. Construct a power grid topology diagram: The nodes are distribution transformer substations, and the edges are line impedances;

[0106] S203-2. Calculate the power fluctuation propagation weights between nodes, and focus on capturing the cascading impact of sudden changes in photovoltaic output on adjacent charging piles;

[0107] Regarding the calculation of the weights:

[0108] First, construct a node association model based on the power grid topology structure, and convert the line impedance between distribution transformer substations into electrical distance weights: The initial weights of adjacent nodes with lower impedance are higher, reflecting the inherent connection strength of the power grid. At the same time, integrate real-time measurement data, analyze the power fluctuation direction and amplitude of each node, and identify the propagation paths of sudden power increases or drops.

[0109] Then, when a power mutation occurs at a certain node, the system automatically tracks its cascading impact on upstream and downstream nodes.

[0110] Method for determining the cascading impact: When a power mutation occurs at a certain node, the system automatically tracks its cascading impact on upstream and downstream nodes - if the voltage fluctuation of the target node shows strong spatio-temporal correlation with the power change of the source node, then enhance the propagation weight between the two nodes; otherwise, reduce the weight. This dynamic adjustment mechanism enables the model to adapt to changes in the power grid operation state and accurately capture the power fluctuation diffusion law under extreme weather conditions such as typhoons.

[0111] Furthermore, design a photovoltaic-charging load cross-attention module using a cross-modal attention mechanism to quantify the mutual interference intensity between the two under extreme weather;

[0112] This module adopts a dual-channel multi-head attention architecture, with the time-series characteristics of photovoltaic output as the query vector and the charging load characteristics as the key-value pair. By calculating the cross-modal attention weights, dynamically capture the guiding effect of photovoltaic fluctuations on charging demand under specific weather conditions (for example, when heavy rain causes a sudden drop in photovoltaic power, the charging demand increases due to user detention). Each attention head focuses on the mutual interference patterns at different time scales, and finally weighted aggregation generates the mutual interference intensity coefficient.

[0113] Quantify the degree of mutual interference: During extreme weather periods (such as when a typhoon passes by), extract the variance of the attention weights as a quantitative index of the mutual interference intensity - the more concentrated the weight distribution, the stronger the causal relationship between photovoltaic and charging load.

[0114] S204. Design the multi-objective loss function of the spatio-temporal association model;

[0115] The design of the multi-objective loss function includes a main loss term and a physical constraint term;

[0116] Further, the main loss term refers to the weighted sum of the mean square errors of photovoltaic and charging power predictions;

[0117] The physical constraint terms include grid power balance constraints, temporal smoothing constraints, and mutual information maximization terms;

[0118] Specifically:

[0119] Grid power balance constraint: Photovoltaic output + charging power ≤ transformer rated capacity;

[0120] Temporal smoothing constraint: Limitation on the jump amplitude of predicted power at adjacent time points;

[0121] Mutual information maximization term: Forcing the model to learn the implicit correlation law between photovoltaic and charging loads;

[0122] Specifically, it can be expressed as:

[0123]

[0124] where L total is the multi-objective loss function; α, β are hyperparameters; L PV is the mean square error of photovoltaic output prediction; L EV is the weighted absolute error of charging power prediction; λ1, λ2 are regularization coefficients; L grid is the loss of grid power balance constraint; L smooth is the temporal smoothness constraint;

[0125] S205. Form a dynamic learning mechanism;

[0126] It includes weather type adaptive learning rate and sliding window incremental training;

[0127] Specifically:

[0128] Weather type adaptive learning rate: The learning rate in rainy scenarios is increased to 1.5 times that in sunny days;

[0129] Sliding window incremental training: Automatically fine-tune the model parameters with new data every 7 days to prevent prediction drift.

[0130] S3. Establish a performance correction mechanism for the intelligent prediction model;

[0131] It includes the following steps:

[0132] S301. Conduct error monitoring and anomaly diagnosis on the intelligent prediction model;

[0133] Specifically, it includes the following steps:

[0134] S301-1. Calculate multi-dimensional error metrics in real time;

[0135] Among them, the multi-dimensional error index includes short-term error, long-term drift, and mutation detection;

[0136] Specifically:

[0137] Short-term error: Rolling calculation of the mean absolute error of photovoltaic and charging power predictions within the latest 15-minute window, and dividing the benchmark threshold according to weather types. For example, on sunny days: ≤5%, and on heavy rain days: ≤12%;

[0138] Long-term drift: Statistical calculation of the cumulative distribution function of prediction errors in the past 72 hours to detect whether the error distribution pattern deviates from the historical pattern;

[0139] Mutation detection: Based on the CUSUM algorithm, identify the sudden deviation between the predicted power and the measured value, and set the mutation event threshold. For example, the error increase exceeds 20% within 1 minute;

[0140] S301-2. Conduct root cause analysis of errors;

[0141] Including: Feature contribution degree backtracking and associated device verification;

[0142] Specifically:

[0143] Feature contribution degree backtracking: Use SHAP values to analyze the contribution intensity of each input feature during the error period and locate the failed features. For example, a failed meteorological sensor causes the irradiance feature to be distorted;

[0144] Associated device verification: When the charging power prediction is abnormal, automatically trigger the BMS self-check instruction of the associated charging pile to exclude prediction deviations caused by device failures;

[0145] S302. Online correction of the intelligent prediction model according to the diagnosis result of S301;

[0146] Specifically, it includes the following steps:

[0147] S302-1. Adopt model hot update: Adopt the elastic weight consolidation algorithm to inject new samples while retaining historical knowledge to prevent catastrophic forgetting;

[0148] S302-2. Conduct feature space optimization;

[0149] Including failed feature masking and derivative feature generation;

[0150] Specifically:

[0151] Failed feature masking: When the SHAP contribution degree of a certain feature is lower than the threshold for 3 consecutive windows, automatically set its weight to zero;

[0152] Derivative feature generation: Real-time calculation of the higher-order statistics of meteorological variables, such as the moving standard deviation of irradiance volatility, and supplement it to the input feature set;

[0153] S302-3. Build a feedback closed-loop;

[0154] It includes a short-term feedback closed-loop and a long-term feedback closed-loop;

[0155] Specifically:

[0156] Short-term feedback: Generate a correction efficiency report every 5 minutes, including the error reduction rate and the feature adjustment list, and push it to the operation and maintenance terminal;

[0157] Long-term feedback: Summarize the correction records daily, optimize the threshold parameters and trigger logic, and conduct strategy simulation through the digital twin platform.

[0158] S4. Based on the intelligent prediction model of the accessory performance correction mechanism, jointly predict the distributed photovoltaic output and the charging station power.

[0159] Embodiment:

[0160] To verify the effectiveness of the present invention, we select a distribution network connected to a photovoltaic-battery-charging station as the research object; the climate feature is subtropical monsoon climate, with an average annual rainfall of 1200 mm; the total installed capacity of distributed photovoltaics is 800 kW, and the component type is monocrystalline silicon PERC components (conversion efficiency 21.6%);

[0161] The types of charging piles include:

[0162] Fast charging piles: 20 sets of 120 kW DC piles, and the charging time from SOC 20% to 80% is 25 minutes;

[0163] Slow charging piles: 50 sets of 7 kW AC piles, with an average daily usage duration of 6 hours;

[0164] As Figure 1 shown is the comparison of photovoltaic output prediction;

[0165] Black solid line: The actual photovoltaic output value. Due to special weather, the power drops suddenly from 13:00 to 15:00, with a decrease of about 70%;

[0166] Red dashed line: The prediction result of the present invention still maintains a high tracking accuracy during the special weather influence period (gray background);

[0167] Blue dotted line: The prediction result of the traditional LSTM fails to respond to sudden weather changes in time, and the maximum error reaches 42%;

[0168] Marked by the red arrow: After detecting the prediction deviation, the present invention triggers the correction mechanism at 13:10, and the prediction curve quickly converges to the actual value;

[0169] Advantage manifestation:

[0170] The prediction error during special weather periods has been reduced from 32% of the traditional method to 9.5%.

[0171] The response delay for power dip events has been shortened to 10 minutes, while the traditional method requires 45 minutes.

[0172] As Figure 2 shown are the charging power prediction and correction effects;

[0173] Green dashed line: The predicted value of the charging power of the present invention, which highly coincides with the true value (black solid line);

[0174] Magenta curve: The original prediction error rate, with an error of up to 18% during the midday peak period (11:00 - 14:00);

[0175] Blue curve: The error rate after dynamic correction, suppressing the error below 9% through online parameter update;

[0176] Orange filled area: The error correction amplitude, with the maximum single - point error decreasing by 62%;

[0177] Advantages are reflected as follows:

[0178] The dynamic correction mechanism reduces the error during peak periods by 50%;

[0179] The average error rate throughout the day has been reduced from 12.3% to 6.8%;

[0180] The error recovery time (ERT) is less than 8 minutes.

[0181] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above - mentioned exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non - restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed within the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

Claims

1. A combined intelligent prediction method for low-voltage distributed photovoltaic output and charging station power, characterized in that, Construct an information feature set between distributed photovoltaic output and user charging willingness under different environmental factors; Establish a joint intelligent prediction model of distributed photovoltaic output and charging station power based on the information feature set and power consumption feature data; Establish a performance correction mechanism for the intelligent prediction model; Carry out the joint prediction of distributed photovoltaic output and charging station power according to the intelligent prediction model with the affiliated performance correction mechanism.

2. The intelligent joint prediction method for low-voltage distributed photovoltaic output and charging station power according to claim 1, wherein The construction of the information feature set includes the following steps: S101. Conduct multi-source data collection and preprocessing; S102. Calculate features jointly with the sensitivity coefficient for the data processed in S101; S103. Establish a feature screening mechanism, screen the features, and obtain the information feature set.

3. The intelligent joint prediction method for low-voltage distributed photovoltaic output and charging station power according to claim 2, characterized in that, In S102, synchronously quantify the dynamic influence of meteorological variables on photovoltaic output and charging behavior through a single coefficient; In S103, the screening of features refers to calculating the maximum information coefficient between each feature and the target variable, retaining the relevant features whose maximum information coefficient exceeds the set threshold, and manually reviewing the rationality of the features.

4. The intelligent joint prediction method for low-voltage distributed photovoltaic output and charging station power according to claim 1, characterized in that The establishment of the intelligent prediction model includes the following steps: S201. Establish an input module for the joint intelligent prediction model of distributed photovoltaic output and charging station power; S202. Establish a feature extraction layer for the joint intelligent prediction model of distributed photovoltaic output and charging station power; S203. Establish a spatio-temporal correlation model between distributed photovoltaic output and charging station power; S204. Design a multi-objective loss function for the spatio-temporal correlation model; S205. Form a dynamic learning mechanism.

5. The joint intelligent prediction method for low-voltage distributed photovoltaic output and charging station power according to claim 4, characterized in that, In step S201, the establishment of the input module includes data channel division and data standardization; Data channel division includes a meteorological-photovoltaic channel and a user-grid channel; Data standardization means implementing maximum-minimum normalization for the data of each channel to eliminate the dimension difference; In S202, the establishment of the feature extraction layer refers to conducting local meteorological feature extraction, including: S202-1. Use a convolutional neural network to capture the local spatial correlation patterns of irradiance and cloud movement; S202-2. Design a 3-layer convolutional structure, and the convolutional kernel size is adaptively adjusted with the time dimension; S202-3. Adopt a bidirectional long short-term memory network to learn the intra-day cycle characteristics of photovoltaic output and charging load, and introduce a weather type gating mechanism; In S203, the establishment of the spatio-temporal correlation model includes: S203-1. Construct a power grid topology map: the nodes are distribution transformer substations, and the edges are line impedances; S203-2. Calculate the power fluctuation propagation weight between nodes to capture the cascading impact of photovoltaic output mutations on adjacent charging piles; In S204, the design of the multi-objective loss function includes a main loss term and a physical constraint term; In S205, the dynamic learning mechanism includes a weather type adaptive learning rate and sliding window incremental training.

6. The intelligent joint prediction method for low-voltage distributed photovoltaic output and charging station power according to claim 1, wherein The establishment of the intelligent prediction model performance correction mechanism includes the following steps: S301. Conduct error monitoring and anomaly diagnosis on the intelligent prediction model; S302. Carry out online correction on the intelligent prediction model according to the diagnosis result of S301.

7. The intelligent joint prediction method for low-voltage distributed photovoltaic output and charging station power according to claim 6, wherein In S301, it includes the following steps: S301-1. Real-time calculation of multi-dimensional error metrics, where the multi-dimensional error metrics include short-term error, long-term drift, and mutation detection; S301-2. Conduct error root cause analysis, including: feature contribution backtracking and associated device verification.

8. The intelligent joint prediction method for low-voltage distributed photovoltaic output and charging station power according to claim 6, characterized in that, In S302, the following steps are included: S302-1. Adopt model hot update: Use the elastic weight consolidation algorithm to inject new samples while retaining historical knowledge; S302-2. Optimize the feature space, including failure feature masking and derivative feature generation; S302-3. Construct feedback loops, including short-term feedback loops and long-term feedback loops.

Images