Distributed photovoltaic short-term prediction method based on artificial intelligence deep spatiotemporal model

Abstract

The present application relates to the technical field of power system data analysis, and particularly relates to a distributed photovoltaic short-term prediction method based on an artificial intelligence deep space-time model. The present application can improve the accuracy of short-term photovoltaic prediction, provide a buffer time for power grid dispatching, adjust the output of adjustable power sources such as thermal power and hydropower, or start the charging and discharging of energy storage systems to suppress power fluctuations, avoid the deviation of power grid frequency and voltage from the rated range, and reduce the risk of equipment damage, line overload and even large-area power outage caused by fluctuations. Precise short-term prediction can help the power grid to evaluate the superposition of photovoltaic output peak and load peak in advance, optimize the dispatching of power transmission channels, maximize the use of existing power grid resources, and thus improve the regional photovoltaic consumption rate.

Description

Technical Field

[0001] This invention relates to the field of power system data analysis technology, and in particular to a method for short-term forecasting of distributed photovoltaic power based on an artificial intelligence deep spatiotemporal model. Background Technology

[0002] In the context of new power system construction, load structures are becoming increasingly complex, and traditional forecasting and analysis methods based on total load are struggling to cope with the challenges brought about by increased load volatility and uncertainty. Existing technologies typically rely on administrative definitions or industry codes to classify loads at a coarse-grained level, failing to reveal the differences in electricity consumption behavior, adjustable potential, and load characteristic evolution mechanisms among different users within the same industry. This results in low load forecasting accuracy and difficulty in supporting refined grid dispatching. Existing forecasting methods have significant limitations in characterizing the spatial relationships of new energy power plants: while CNN-LSTM-based convolutional-temporal fusion models use matrix encoding to represent power plant locations, this simplified processing only retains basic directional relationships, completely losing crucial spatial distance information, leading to serious defects in spatial feature representation; while graph neural network-based methods can construct site association networks, their node connection rules heavily rely on empirical threshold settings. This highly subjective approach is ill-suited to the diverse power plant layout characteristics of different regions. Neither of these mainstream methods can accurately capture the true spatial dependencies between power plants, severely limiting the adaptability of forecasting models to regional meteorological propagation patterns. To address this critical issue, this project innovatively abandoned the traditional matrix or graph structure encoding mode and constructed a new spatial representation paradigm that can simultaneously preserve the location relationship of power stations and accurate distance information. This enabled the accurate depiction of the regional spatial topology and laid a solid foundation for subsequent deep feature extraction.

[0003] Furthermore, the biggest drawback of photovoltaic (PV) power plants lies in the severe energy efficiency mismatch between the power grid and end users. Power generation and load are often measured separately, and the results typically only indicate whether the PV power plant is suitable for the end user, failing to effectively match applications. In particular, nighttime PV curtailment often mismatches with peak demand, leading to a direct abandonment of power generation to match load. This results in energy being either stored or largely wasted. Storing energy involves unnecessary energy conversion, and the associated costs contribute to even greater waste. How to quickly and effectively achieve short-term PV forecasting and adapt it to different companies remains a key technical challenge. Summary of the Invention

[0004] This invention aims to significantly improve the accuracy and reliability of short-term photovoltaic power forecasting, while enhancing the transparency of model decision-making, thus providing strong technical support for the safe and stable operation of the power grid and the efficient absorption of new energy sources. This addresses the aforementioned technical problems.

[0005] A method for short-term forecasting of distributed photovoltaic power based on an artificial intelligence deep spatiotemporal model includes the following steps: S1. Acquire historical data of multiple photovoltaic power stations in the region. The historical data includes latitude and longitude, photovoltaic output, irradiance, and temperature. The historical data is preprocessed to form multi-channel spatial image encoding and generate multi-channel spatiotemporal tensors. S2. Input the multi-channel spatiotemporal tensor into the adaptive weighted spatial feature fusion network to obtain the fused feature tensor; S3. Perform gradient-weighted spatial saliency mapping on the fused feature tensor to generate a spatial feature saliency heatmap; S4. Input the spatially significant heatmap into the Conv-ConvLSTM prediction model to output the short-term photovoltaic power output prediction results.

[0006] The multi-channel spatial image encoding in S1 is processed by the multi-channel spatial image encoding module. A two-dimensional spatial grid is constructed based on the latitude and longitude of the power station. Min-Max normalization is used to [0,1], and photovoltaic output, irradiance, and temperature are mapped to different channel pixels respectively. The orientation relationship and accurate geographical distance between power stations are calculated and preserved. The adaptive weighted spatial feature fusion network in S2 processes the three independent convolutional branches of photovoltaic power output, irradiance, and temperature through an adaptive weighted feature fusion module, extracting high-dimensional features respectively; it learns a trainable spatial weight matrix for each branch, and achieves adaptive weighted fusion through Hadamard product.

[0007] In S3, the gradient-weighted spatial saliency mapping is processed by the spatial saliency mapping module to calculate the gradient tensor of the predicted output with respect to the feature map of the last convolutional layer; the channel saliency weights are obtained through global average pooling; and the weighted combination generates a visual heatmap to locate the spatial regions that play a key role in the prediction.

[0008] In S4, a Conv-ConvLSTM prediction model is used to generate deep spatial features of saliency mapping by generating multi-layer two-dimensional convolution; the ConvLSTM layer synchronously models the temporal dynamic evolution and spatial propagation law; and the photovoltaic power output prediction value is output after global average pooling and fully connected layer.

[0009] S5. Collect multi-source data from the load side, meteorological side, and auxiliary information side, and complete anomaly detection, hierarchical repair, time alignment, and desensitization preprocessing. S6. Based on industry attributes, electricity consumption behavior, and dispatch characteristics, perform refined load classification, standardize user daily load curves, extract features and reduce dimensionality, use improved K-means++ clustering to extract typical daily load curves, and obtain classified load aggregation curves through virtual aggregation; S7. Conduct seasonal, weekly, and daily time-domain characteristic analysis on the classified loads, calculate stability and predictability, and complete the classification. S8. Based on the sliding window significance test, quantify the correlation, instantaneous sensitivity and fluctuation contribution of classified load and total load; S9. Construct customized prediction models according to stability predictability classification to obtain load prediction results for each category; S10. The classification prediction results are optimally synthesized using dynamic weighting or cointegration theory, and the total load prediction value of the system is output. The photovoltaic output prediction value of S4 and the total load prediction value of S10 are then matched.

[0010] The multi-source data in S5 is taken from photovoltaic power plants. The collected data includes daily maximum load, daily minimum load, daily peak-to-valley difference, peak-to-valley difference rate, daily average load, daily load rate, daily load fluctuation coefficient, daily peak hours, and daily median load. The collection time granularity is 15 minutes, with 96 collection points in 24 hours. Meteorological side: meteorological forecast data, including temperature, humidity, wind speed, wind direction, precipitation, and air pressure. Auxiliary information side: structured information such as holidays, special events, planned maintenance, and load transfer records. The refined load classification in S6 includes: agricultural irrigation and drainage, power and heat production and supply, steel processing and smelting, silicon carbide, cement manufacturing, and other industries.

[0011] The extraction of the daily load curve in S6 includes: normalizing the time-domain features of days, months, and years, using t-SNE dimensionality reduction to improve K-means++ clustering, determining the optimal number of clusters through the silhouette coefficient, and extracting typical curves according to the proportion of days.

[0012] The S8 calculation of stability and predictability and the subsequent classification are specifically obtained through a one-sided forward difference approximation in a 15-minute discrete sampling environment under continuous-time limits. Let the sampling interval Δt = 15 minutes, then the instantaneous elasticity is εᵢ. ,t ≈(ΔP total,t / P total,t ]÷(ΔP i,t / P i,t ) Where Δ t =15 minutes, P total,t For the total load, ΔP total,t For changes in total load, P i,t For the i-th type of load, ΔP i,t Classify the load change as category i; In S9, load stability is measured by the load variation coefficient, i.e. Where σ is the standard deviation of the load sequence of load i at time t, and μ is the standard deviation of the load sequence of load i at time t; the larger the value, the higher the load stability; the smaller the value, the more susceptible the load is to the influence of the load side, the meteorological side, and the auxiliary information side. The hierarchical prediction model is as follows: The PCA-SARIMAX model is used for highly stable and predictable loads; The stable and predictable load adopts the Stacking ensemble learning model; The low-stability, predictable load adopts the ConvLSTM-Transformer two-path spatiotemporal model.

[0013] In S10, the optimal synthesis adopts dynamic weighted ensemble learning, which uses classification load contribution and sensitivity as weights for weighted fusion.

[0014] The beneficial effects of this invention are: This application improves the accuracy of short-term photovoltaic (PV) forecasts, providing a buffer time for grid dispatch. By adjusting the output of adjustable power sources such as thermal and hydropower, or by activating energy storage systems for charging and discharging, it smooths out power fluctuations, preventing grid frequency and voltage from deviating from their rated ranges, and reducing the risk of equipment damage, line overload, or even large-scale power outages caused by fluctuations. Accurate short-term forecasts can help the grid assess the overlap of peak PV output and peak load in advance, optimize transmission channel dispatch, maximize the utilization of existing grid resources, and thus improve the regional PV integration rate. Detailed Implementation

[0015] A method for short-term forecasting of distributed photovoltaic power based on an artificial intelligence deep spatiotemporal model includes the following steps: S1. Acquire historical data of multiple photovoltaic power stations in the region. The historical data includes latitude and longitude, photovoltaic output, irradiance, and temperature. The historical data is preprocessed to form multi-channel spatial image encoding and generate multi-channel spatiotemporal tensors. S2. Input the multi-channel spatiotemporal tensor into the adaptive weighted spatial feature fusion network to obtain the fused feature tensor; S3. Perform gradient-weighted spatial saliency mapping on the fused feature tensor to generate a spatial feature saliency heatmap; S4. Input the spatially significant heatmap into the Conv-ConvLSTM prediction model to output the short-term photovoltaic power output prediction results.

[0016] The multi-channel spatial image encoding in S1 is processed by the multi-channel spatial image encoding module. A two-dimensional spatial grid is constructed based on the latitude and longitude of the power station. Min-Max normalization is used to [0,1], and photovoltaic output, irradiance, and temperature are mapped to different channel pixels respectively. The orientation relationship and accurate geographical distance between power stations are calculated and preserved. The adaptive weighted spatial feature fusion network in S2 processes the three independent convolutional branches of photovoltaic power output, irradiance, and temperature through an adaptive weighted feature fusion module, extracting high-dimensional features respectively; it learns a trainable spatial weight matrix for each branch, and achieves adaptive weighted fusion through Hadamard product.

[0017] In S3, the gradient-weighted spatial saliency mapping is processed by the spatial saliency mapping module to calculate the gradient tensor of the predicted output with respect to the feature map of the last convolutional layer; the channel saliency weights are obtained through global average pooling; and the weighted combination generates a visual heatmap to locate the spatial regions that play a key role in the prediction.

[0018] In S4, a Conv-ConvLSTM prediction model is used to generate deep spatial features of saliency mapping by generating multi-layer two-dimensional convolution; the ConvLSTM layer synchronously models the temporal dynamic evolution and spatial propagation law; and the photovoltaic power output prediction value is output after global average pooling and fully connected layer.

[0019] It also includes steps S5. Collect multi-source data from the load side, meteorological side, and auxiliary information side, and complete anomaly detection, hierarchical repair, time alignment, and desensitization preprocessing. S6. Based on industry attributes, electricity consumption behavior, and dispatch characteristics, perform refined load classification, standardize user daily load curves, extract features and reduce dimensionality, use improved K-means++ clustering to extract typical daily load curves, and obtain classified load aggregation curves through virtual aggregation; S7. Conduct seasonal, weekly, and daily time-domain characteristic analysis on the classified loads, calculate stability and predictability, and complete the classification. S8. Based on the sliding window significance test, quantify the correlation, instantaneous sensitivity and fluctuation contribution of classified load and total load; S9. Construct customized prediction models according to stability predictability classification to obtain load prediction results for each category; S10. The classification prediction results are optimally synthesized using dynamic weighting or cointegration theory, and the total load prediction value of the system is output. The photovoltaic output prediction value of S4 and the total load prediction value of S10 are then matched.

[0020] The multi-source data in S5 is taken from photovoltaic power plants. The collected data includes daily maximum load, daily minimum load, daily peak-to-valley difference, peak-to-valley difference rate, daily average load, daily load rate, daily load fluctuation coefficient, daily peak hours, and daily median load. The collection time granularity is 15 minutes, with 96 collection points in 24 hours. Meteorological side: meteorological forecast data, including temperature, humidity, wind speed, wind direction, precipitation, and air pressure. Auxiliary information side: structured information such as holidays, special events, planned maintenance, and load transfer records. The refined load classification in S6 includes: agricultural irrigation and drainage, power and heat production and supply, steel processing and smelting, silicon carbide, cement manufacturing, and other industries.

[0021] The extraction of the daily load curve in S6 includes: normalizing the time-domain features of days, months, and years, using t-SNE dimensionality reduction to improve K-means++ clustering, determining the optimal number of clusters through the silhouette coefficient, and extracting typical curves according to the proportion of days.

[0022] The S8 calculation of stability and predictability and the subsequent classification are specifically obtained through a one-sided forward difference approximation in a 15-minute discrete sampling environment under continuous-time limits. Let the sampling interval Δt = 15 minutes, then the instantaneous elasticity is εᵢ. ,t ≈(ΔP total,t / P total,t ]÷(ΔP i,t / P i,t ) Where Δ t =15 minutes, P total,t For the total load, ΔP total,t For changes in total load, P i,t For the i-th type of load, ΔP i,t Classify the load change as category i; In S9, load stability is measured by the load variation coefficient, i.e. Where σ is the standard deviation of the load sequence of load i at time t, and μ is the standard deviation of the load sequence of load i at time t; the larger the value, the higher the load stability; the smaller the value, the more susceptible the load is to the influence of the load side, the meteorological side, and the auxiliary information side. The hierarchical prediction model is as follows: The PCA-SARIMAX model is used for highly stable and predictable loads; The stable and predictable load adopts the Stacking ensemble learning model; The low-stability, predictable load adopts the ConvLSTM-Transformer two-path spatiotemporal model.

[0023] In S10, the optimal synthesis adopts dynamic weighted ensemble learning, which uses classification load contribution and sensitivity as weights for weighted fusion.

[0024] This application delves into six key areas: integrating photovoltaic power generation with large customers; constructing a smart algorithm library for load forecasting categorized by dispatch; monitoring and dispatching response mechanisms for load fluctuations of key large users; assessing the dispatching impact of load characteristics of key large users; and applying the improved accuracy of system total load forecasting considering refined load categorization in dispatching. The aim is to establish a precise, intelligent, efficient, and user-friendly load forecasting method, providing reliable scientific support for key operations such as regional power grid power balance and power generation planning; providing technical support for short-term system load forecasting of regional power grids, significantly improving forecasting accuracy; and providing auxiliary decision-making basis for regional power grids to adapt to the requirements of power market reform and ensure the safe, stable, and economical operation of the power grid in the power market environment. This will promote a steady increase in the company's economic efficiency, performance evaluation, and load control efficiency, enhancing the effectiveness of digital transformation. By adjusting the output of adjustable power sources such as photovoltaic, thermal, and hydropower, or by activating energy storage systems for charging and discharging, power fluctuations can be smoothed, preventing grid frequency and voltage deviations from rated ranges and reducing the risk of equipment damage, line overload, or even large-scale power outages caused by fluctuations.

[0025] 1. This research breaks through the limitations of traditional load analysis, which focuses solely on total load and time series, and constructs a multi-dimensional load characteristic analysis system encompassing spatiotemporal, meteorological, economic, and social activities. It innovatively introduces high-frequency data acquisition and fusion technology, deeply correlating load data with refined weather forecasts (such as perceived temperature), holiday types, major event signals, and regional economic indicators. Through the development of a multi-dimensional coupled analysis model, the system systematically reveals the quantitative impact mechanism of different external factors on the shape, peak value, and variation patterns of the load curve. This research provides a new analytical perspective and a solid theoretical foundation for accurately grasping the driving forces of load changes and identifying potential growth points and sensitive factors, significantly enhancing the depth and breadth of our understanding of load characteristics.

[0026] 2. This application focuses on key industries that dominate regional load, conducting refined and in-depth load characteristic profiling. Its innovation lies not only in depicting typical daily, weekly, and yearly load curves, but also in deeply analyzing the causal relationship between their production processes, production scheduling plans, business models, and other intrinsic drivers and load demand. By connecting with actual enterprise operating data, it reveals the composition ratio of "flexible load" and "rigid load" within the industry, as well as the distribution of adjustable potential. This work represents a leap from "phenomenon description" to "mechanism analysis," laying a crucial data and model foundation for subsequent industry-specific precise forecasting and the exploration of demand-side response resources.

[0027] 3. Addressing the bottlenecks of traditional methods for extracting typical load curves, which rely on human experience, are highly subjective, and struggle with massive user data, this project innovatively applies and improves upon various unsupervised clustering algorithms (such as K-means, DBSCAN, and hierarchical clustering). It focuses on solving the problems of effective dimensionality reduction and feature extraction for high-dimensional and nonlinear load data, and automatically optimizes the number of clusters using metrics such as the silhouette coefficient. This method can automatically and objectively identify representative typical load patterns from massive user data, effectively removing outliers and noise interference, greatly improving the efficiency and accuracy of typical curve extraction. It provides clean and reliable input for load characteristic classification research and subsequent industry-specific and type-specific prediction model construction.

[0028] 4. This application delves into the inherent laws governing the complexity and variability of load characteristics across different industries. It innovatively constructs an evaluation index system for industry load complexity, including dimensions such as volatility, periodicity, randomness, and sensitivity to external factors. The application systematically studies the essential differences in complexity among different industries (e.g., continuous production in industry and intermittent operation in commerce) and their variability evolving with seasons, policies, and market environments. By introducing complexity science theory, it reveals the critical conditions for stable and abrupt changes in load characteristics, deepening the understanding of the sources of load uncertainty. This research provides a theoretical basis for assessing forecasting difficulty, developing differentiated forecasting strategies, and improving the system's ability to cope with load abrupt change risks.

[0029] The above technical solution is further explained as follows: Load side: Covering all 10kV and above dedicated transformer users and public transformer substations in the power grid area, the collected data includes daily maximum load, daily minimum load, daily peak-to-valley difference, peak-to-valley difference rate, daily average load, daily load factor, daily load fluctuation coefficient, daily peak hours, and daily median load. The data collection time granularity is 15 minutes.

[0030] Daily maximum load P max : Maximum load value in the 24 hours of the day.

[0031] Daily minimum load P min Minimum load value in the 24 hours of the day.

[0032] Daily peak-to-valley difference ΔP: The difference between the maximum and minimum load values ​​for the day, i.e., ΔP = P max -P min .

[0033] Peak-to-valley difference rate ΔP%: The percentage of peak-to-valley difference relative to the maximum load, i.e. .

[0034] Daily average load Pavg: The average load over the 24 hours of the day.

[0035] Daily load factor (LF): The ratio of the average daily load to the maximum daily load. .

[0036] Daily load fluctuation coefficient (CV): the ratio of the standard deviation of the load to the average load, i.e. The smaller the dimensionless fluctuation intensity (CV), the smoother the curve and the narrower the prediction error range.

[0037] Peak Hours: Daily load ≥ 90% max The number of hours.

[0038] Daily load median P50: The value that is in the middle after the 24-hour load is sorted by size.

[0039] Meteorological side: Meteorological forecast data, including temperature, humidity, wind speed, wind direction, precipitation, and air pressure.

[0040] Auxiliary information side: structured information such as holidays, special events, planned maintenance, and load transfer records.

[0041] Holidays: Statutory holidays and work adjustment arrangements announced by the state (such as Spring Festival, National Day, etc.).

[0042] Special events: Unplanned events that occur temporarily at the government or social level and have a significant impact on electricity consumption behavior, including large-scale sporting events, concerts, extreme weather emergencies, and public health emergencies.

[0043] Planned maintenance: Power grid outage maintenance plan. Load transfer record: Power grid load transfer dispatch record.

[0044] The collected data is load operation data, forming a multi-dimensional information closed loop by collecting load-side, meteorological, and auxiliary information: on the load side, electricity and load data are continuously captured; on the meteorological side, meteorological data is captured; and on the auxiliary information side, information on holidays, special events, planned maintenance, and load transfer records is updated synchronously.

[0045] The following issues exist in the load operation data: (1) Null value: The entire line or multiple consecutive points are missing due to power failure of the terminal or communication interruption; (2) Zero value / negative value: Non-physical result caused by abnormal meter reading return code or wiring error; (3) Continuous and constant: A "straight line" for a long time usually indicates that the terminal has crashed or frozen; (4) Step change: a jump on the second or minute level, commonly seen in load switching and switch maintenance; (5) Outlier peaks: Single point or local extreme values, often caused by harmonic impulses or instantaneous short circuits; (6) Cross-system misalignment: "pseudo-anomalies" caused by timestamp drift, dimensional differences, and holiday misalignment.

[0046] Power system data cleaning removes erroneous, duplicate, and incomplete data, improving its accuracy and completeness, thus providing reliable data support for power system operation, management, analysis, and decision-making. Using uncleaned data directly will severely impact the accuracy of subsequent processes such as load characteristic identification, cluster analysis, correlation modeling, and algorithm training. Cleaned data effectively reduces false alarms and missed alarms, optimizes power system operating efficiency, lowers costs, supports the application of smart grid technologies, and ensures power system security. Furthermore, high-quality data facilitates data sharing and long-term management among different departments, laying a solid foundation for the intelligent and sustainable development of the power system. Chebyshev's inequality, as a classic probabilistic tool, requires no prior knowledge of the distribution of random variables. This property allows it to maintain robust theoretical guarantees even when dealing with load curves containing periodic, peak, and heavy-tailed characteristics. By setting confidence coefficients, abnormal observations far from the mean can be quickly identified with an upper bound probability not exceeding 1 / k² without relying on distribution assumptions.

[0047] The project uses 96 points (with data points extracted every 15 minutes within 24 hours) as a sliding window to estimate the mean μ and standard deviation σ segment by segment. This process is deployed simultaneously in both Flink real-time streaming and Spark offline batch modes. The latency of the former is controlled in the second level, while the latency of the latter is controlled in the minute level for daily full scans, ensuring timeliness and scalability in large-scale data scenarios.

[0048] For the abnormal samples and missing data identified in the above steps, a hierarchical repair framework is established to ensure that the repaired curves maintain minimal deviation from the original values ​​in terms of statistical characteristics, physical rationality, and business semantics.

[0049] The three basic principles of the repair framework are: Minimal intervention principle: Repair is only performed on confirmed abnormal intervals, while preserving the details of normal fluctuations; Physical consistency principle: Repair results must meet electrical constraints such as contract capacity, power factor, and energy conservation; Business traceability: A repair log is generated for each repair segment, recording the method, basis, confidence interval, and responsible person.

[0050] A layered repair framework is adopted based on the size of the gap or abnormal region: 1. Short-term layer: Minor repair (<1h) Common causes: intermittent communication interruptions or sporadic data loss due to terminal buffer overflow.

[0051] Method: Linear interpolation.

[0052] S1. Take 4 points on each side of the gap as interpolation nodes; S2. Calculate the first derivative of the node as the boundary slope to prevent Runge oscillation; S3. Generate a third-order polynomial and output the 4-point interpolation result; The boundary slope limit is ≤ 1.5 times the average slope of adjacent segments.

[0053] 2. Mid-time zone: Repair work (1–4 hours) Common causes: Power outage during pre-planned maintenance testing, or power supply reversal in some transformer substations.

[0054] Method: Weighted reconstruction based on the same day.

[0055] S1. Feature retrieval: Weekday type, temperature range (±2℃), and event tag 3D key value; S2. Similarity Calculation: Calculate the DTW distance for the 96-point curve after z-score normalization. Find the minimum distance curve for k=5; S3. Weight Allocation: , Take the mean of the five curves DTW; S4. Weighted Fusion: Generate Repair Segments ; S5. Cosine window smoothing: Apply a window function transition 2 hours before and after the notch to suppress splicing marks; Note: parameter: Adaptive updates, recalculated monthly; 3. Long-duration layer: generated (≥4 hours or entire dead zone) Common causes: prolonged power outage in the distribution area, or batch failure of terminals.

[0056] Method: Multivariate LSTM-GAN generative completion.

[0057] S1. Data Tensor: 96-point power sequence + 5-dimensional meteorological data + 3-dimensional special event generation tensor data; S2. Network architecture: LSTM encoder (128) → Latent (64) → LSTM decoder (128) → output layer; discriminator is 1-DCNN (64, 3×3); S3. The loss function is adjusted using weights; S4. Training: Incremental training every quarter, Adam optimizer, learning rate 1e-3, batch size 256; S5. Post-processing: If the predicted mean is greater than the contract capacity × 1.05, perform linear compression and record the compression coefficient α; S6. Confidence Interval: Generates 90th percentile curves for subsequent uncertainty weighting in prediction; Parameters: Hidden dimension 64, training epoch 100; The load stability is shown in Table 1. A comprehensive analysis of the load stability of agricultural irrigation and drainage, electricity and heat production and supply, steel processing and smelting, silicon carbide, cement manufacturing, other industrial users, and total load reveals that differences in load stability mainly stem from factors such as production continuity, changes in market demand, stability of raw material supply, seasonal demand fluctuations, policy adjustments (including environmental and electricity pricing policies), and holiday effects. Industries with strong production continuity, such as electricity and heat production, steel processing and smelting, silicon carbide, and cement manufacturing, exhibit relatively stable loads. In contrast, agricultural irrigation and drainage and other industrial users are significantly affected by weather conditions, seasonal changes, market demand, and policy adjustments, resulting in relatively lower load stability. Furthermore, electricity price fluctuations also significantly impact the production plans and load stability of energy-intensive industries such as silicon carbide. These common patterns provide important reference value for grid operators in load forecasting, grid planning, and demand-side management.

[0058] Predictability rating measures the degree to which different types of loads can be accurately predicted within a specific time scale. This project, through comprehensive analysis of historical load data, periodic characteristics, sensitivity to exogenous factors (such as meteorological conditions), and the decay trend of historical prediction errors, classifies the predictability of seven types of loads into three levels: high, medium, and low.

[0059] I. Predictability Grading Criteria 1. High predictability: Characteristics: The load curve exhibits a clear periodicity, is less affected by exogenous factors (such as weather conditions and holidays), and historical prediction errors decay rapidly.

[0060] Applicable algorithms: Traditional statistical models (such as time series analysis) and simple machine learning models (such as linear regression) can achieve good prediction results.

[0061] Application scenarios: Suitable for scenarios with high prediction accuracy requirements and limited computing resources.

[0062] 2. Medium Predictability: Characteristics: The load curve has a certain periodicity, but it is greatly affected by exogenous factors, and the historical prediction error decays at a moderate rate.

[0063] Applicable algorithms: Machine learning models that require the integration of exogenous factors (such as meteorological data and holiday information) (such as random forests and gradient boosting trees) or hybrid models (such as Prophet).

[0064] 3. Application Scenarios: Suitable for applications requiring high prediction accuracy and with sufficient computing resources. The scenario is characterized by low predictability. Characteristics: The load curve has no obvious periodicity, is greatly affected by exogenous factors, and the historical prediction error decays slowly.

[0065] Applicable algorithms: Require the use of complex deep learning models (such as LSTM, Transformer) to capture complex nonlinear relationships and temporal dependencies.

[0066] Application scenarios: Suitable for scenarios with extremely high prediction accuracy requirements and sufficient computing resources.

[0067] II. Load Predictability Analysis 1. Agricultural irrigation and drainage Predictability: Low Reasons: Agricultural irrigation and drainage loads are affected by a variety of factors, including weather changes (such as rainfall, etc.). Factors such as drought, seasonal demand (crop growth cycle), and government policies (water resource management, irrigation subsidies) all increase the uncertainty and volatility of loads. Furthermore, soil types, crop varieties, and irrigation technologies in different regions also lead to load variations, making forecasting even more complex.

[0068] Applicable algorithms: These require complex machine learning models (such as random forests and gradient boosters) or deep learning models (such as LSTM networks) to capture the impact of these complex factors on the load.

[0069] Practical application: In practical applications, it is necessary to combine multi-source information such as meteorological data, soil moisture data and crop growth data for prediction, and continuously adjust and optimize the model according to the actual situation.

[0070] 2. Electricity and heat production and supply industry Predictability: High Reason: This type of load has extremely high stability, its production process is continuous and highly regular, and it is minimally affected by exogenous factors (such as weather conditions). The load curve almost presents a high-level plateau line, exhibiting a clear periodic pattern.

[0071] Applicable algorithms: Traditional statistical models such as time series analysis (e.g., SARIMA) and linear regression can achieve good prediction results.

[0072] Practical applications: Due to its high predictability, a lightweight prediction model can be used to reduce the consumption of computing resources, and the prediction model can be updated in a timely manner when production plans are adjusted.

[0073] 3. Iron and steel processing and smelting Predictability: Medium Reasons: Although the production process is relatively continuous, the load may fluctuate due to factors such as raw material supply, market demand, environmental policies, and electricity price fluctuations. Furthermore, equipment maintenance and repairs during the production process can also affect load stability.

[0074] Applicable algorithms: These require a combination of time series analysis and machine learning models (such as support vector machines and decision trees) to predict loads.

[0075] Practical application: In practical applications, it is necessary to consider adjustments to production plans and equipment maintenance plans, and update the forecasting model in a timely manner to reflect these changes.

[0076] 4. Silicon carbide Predictability: Medium Reason: The production process is continuous, but it is greatly affected by fluctuations in electricity prices and market demand. Enterprises may adjust their production plans based on electricity prices, leading to load fluctuations.

[0077] Applicable algorithms: Require the use of machine learning models (such as random forests and gradient boosters) to predict loads.

[0078] Practical application: In practical applications, it is necessary to combine electricity price information and market demand data for forecasting, and continuously adjust and optimize the model according to the actual situation.

[0079] 5. Cement manufacturing Predictability: Medium Reasons: The production process is relatively continuous, but the load may fluctuate due to environmental policies, market demand, and off-peak production policies. In addition, equipment maintenance and repair during the production process can also affect the stability of the load.

[0080] Applicable algorithms: Require a combination of time series analysis and machine learning models (such as support vectors). (Machine, decision tree) to predict load.

[0081] Practical application: In practical applications, it is necessary to consider adjustments to production plans and equipment maintenance plans, and update the forecasting model in a timely manner to reflect these changes.

[0082] 6. Other industrial users Predictability: High Reason: Production plans are relatively fixed and load stability is good, but load fluctuations may occur due to market demand, raw material supply, environmental policies and holidays.

[0083] Applicable algorithms: Time series analysis (such as ARIMA) and machine learning models (such as decision trees and neural networks) can be used to predict loads.

[0084] Practical application: In practical applications, it is necessary to consider changes in market demand and the holiday effect, and update the forecasting model in a timely manner to reflect these changes.

[0085] Through the analysis of load stability and predictability, this project provides scientific management strategies and algorithm selection for power grid dispatching and load forecasting. Different load types with varying stability and predictability require different forecasting models and strategies to optimize the allocation of computing resources, improve forecasting accuracy, and provide strong support for the stable operation and economical dispatch of the power grid.

[0086] In load analysis, stability and predictability are two closely related dimensions with an inherent logical connection. Stability refers to the regularity and repetitiveness of the load curve over time, while predictability refers to the ability to accurately predict future loads based on historical data and exogenous factors. A highly stable load curve typically exhibits significant periodicity and low random fluctuations; this regularity makes the load curve highly predictable in the future.

[0087] 1. Highly stable and highly predictable load Management strategy: Applicable to the "lightweight model + real-time correction" strategy. Using SARIMA as the baseline model, and combining production plans and equipment maintenance information, abnormal events are manually corrected.

[0088] This strategy has low computational resource consumption, is suitable for deployment on edge nodes, and can quickly respond to production needs. Adjustments to the plan.

[0089] Algorithm selection: SARIMA was used as the baseline model, and adjustments were made based on production planning information.

[0090] In the event of an anomaly, real-time corrections are made through manual intervention to ensure prediction accuracy.

[0091] Application scenarios: Suitable for scenarios with high prediction accuracy requirements and limited computing resources, such as... Basic load forecasting for the power grid.

[0092] 2. Medium-stable to medium-predictable load Management strategy: Multi-dimensional feature engineering needs to be introduced. For these load types, the focus should be on adding the original... Factors influencing this include material supply, market demand, and environmental policies. A hybrid model, such as Stacking, is employed. The method combines statistical models with machine learning models, which can significantly reduce prediction errors.

[0093] Algorithm selection: Random forest or gradient boosting tree models that combine raw material supply, market demand, and environmental policies.

[0094] Application scenarios: Suitable for scenarios with high requirements for prediction accuracy and sufficient computing resources, such as load forecasting in industrial areas.

[0095] 3. Low stability - low predictability load Management Strategy: A spatiotemporal deep learning framework is recommended. Construct a "weather-irrigation demand" attention module to capture sudden changes under extreme weather conditions.

[0096] Algorithm selection: LSTM or Transformer model that combines meteorological data and agricultural activity information.

[0097] Application scenarios: Suitable for scenarios with extremely high prediction accuracy requirements and sufficient computing resources, such as load prediction in agricultural irrigation areas.

[0098] This application summarizes the proposed data acquisition and preprocessing workflow. Data acquisition covers three categories: load side, climate side, and auxiliary information, providing comprehensive data support for subsequent analysis. In the data cleaning stage, Chebyshev's inequality is used to identify abnormal data, and a layered repair framework is adopted based on the size of gaps or abnormal intervals to ensure data quality. For data alignment, a triple verification mechanism of "clock-value-semantic" is constructed to ensure data consistency. For data anonymization, a three-step process of hierarchical labeling, strategy orchestration, and audit loop is adopted to ensure data privacy and compliance, while balancing data accuracy and privacy protection. The application also provides an in-depth analysis of the load composition and characteristics of the municipal power grid. By comprehensively considering industry attributes, similarity of electricity consumption behavior, and the feasibility of regulation, the load is divided into six categories. This classification system covers diverse electricity consumption scenarios such as agricultural irrigation and drainage, steel, and cement, laying a solid foundation for subsequent refined analysis. Based on the classification, the study further extracts typical daily load curves and uses virtual aggregation technology to integrate similar electricity consumption patterns. This process not only accurately characterizes the features of each load component but also provides a clear framework for subsequent analysis. Through in-depth analysis of typical daily load curves, the study reveals the electricity consumption patterns and fluctuation laws of different load components within a typical day, providing important data support for the construction of subsequent prediction models.

[0099] In the overall load characteristic analysis phase, the study delves into the load characteristics from three key dimensions: time-domain characteristics, stability, and predictability. It analyzes the load characteristics of various load categories at different time scales. The research not only reveals the inherent patterns of the load but also provides crucial evidence for constructing prediction models. Specifically, the time-domain characteristic analysis covers seasonal latitude, intra-weekly latitude, and intra-day latitude. By analyzing load characteristics, the study accurately captures the fluctuation patterns of different load categories.

[0100] Furthermore, the greatest innovation of this application lies in the two-way assessment of power supply and consumption, integrating photovoltaic power generation with thermal and hydropower, and forming an interactive matching with power consumers. This enables networked intelligent allocation of photovoltaic power stations in different locations, significantly expanding photovoltaic application scenarios. It interconnects photovoltaic power supply with high-power-consuming units such as agricultural irrigation and drainage, electricity and heat production and supply, steel processing and smelting, silicon carbide, and cement manufacturing. While existing technologies mostly analyze power supply and load, this application focuses on analyzing multi-source data taken from photovoltaic power stations. The collected data includes daily maximum load, daily minimum load, daily peak-to-valley difference, peak-to-valley difference rate, daily average load, daily load factor, daily load fluctuation coefficient, daily peak hours, and daily median load, achieving 24-hour effective monitoring and effectively avoiding the technical problem of insufficient power supply at night. By reducing grid fluctuations through stability and predictability, it provides a highly stable load curve with obvious periodicity and low random fluctuation. This regularity makes the load curve highly predictable in the future.

Claims

1. A distributed photovoltaic short-term forecasting method based on an artificial intelligence deep spatiotemporal model, characterized in that... Includes the following steps: S1. Acquire historical data of multiple photovoltaic power stations in the region. The historical data includes latitude and longitude, photovoltaic output, irradiance, and temperature. The historical data is preprocessed to form multi-channel spatial image encoding and generate multi-channel spatiotemporal tensors. S2. Input the multi-channel spatiotemporal tensor into the adaptive weighted spatial feature fusion network to obtain the fused feature tensor; S3. Perform gradient-weighted spatial saliency mapping on the fused feature tensor to generate a spatial feature saliency heatmap; S4. Input the spatially significant heatmap into the Conv-ConvLSTM prediction model to output the short-term photovoltaic power output prediction results.

2. The distributed photovoltaic short-term forecasting method based on an artificial intelligence deep spatiotemporal model according to claim 1, characterized in that, The multi-channel spatial image encoding in S1 is processed by the multi-channel spatial image encoding module. A two-dimensional spatial grid is constructed based on the latitude and longitude of the power station. Min-Max normalization is used to [0,1], and photovoltaic output, irradiance, and temperature are mapped to different channel pixels respectively. The orientation relationship and accurate geographical distance between power stations are calculated and preserved. In S2, the adaptive weighted spatial feature fusion network processes the three independent convolutional branches of photovoltaic power output, irradiance, and temperature to extract high-dimensional features respectively. It learns trainable spatial weight matrices for each branch and achieves adaptive weighted fusion through Hadamard product. In S3, the gradient-weighted spatial saliency mapping is processed by the spatial saliency mapping module to calculate the gradient tensor of the predicted output with respect to the final convolutional feature map. Global average pooling is used to obtain the channel saliency weights. Weighted combinations generate a visual heatmap to locate the spatial regions that play a key role in the prediction. In S4, a Conv-ConvLSTM prediction model is used to generate multi-layer two-dimensional convolutions to extract deep spatial features from the saliency mapping. The ConvLSTM layers synchronously model the temporal dynamic evolution and spatial propagation laws. Global average pooling and fully connected layers output the predicted photovoltaic power output value.

3. The distributed photovoltaic short-term forecasting method based on an artificial intelligence deep spatiotemporal model according to claim 1, characterized in that, Its features It also includes steps S5. Collect multi-source data from the load side, meteorological side, and auxiliary information side, and complete anomaly detection, hierarchical repair, time alignment, and desensitization preprocessing. S6. Based on industry attributes, electricity consumption behavior, and dispatch characteristics, perform refined load classification, standardize user daily load curves, extract features and reduce dimensionality, use improved K-means++ clustering to extract typical daily load curves, and obtain classified load aggregation curves through virtual aggregation; S7. Conduct seasonal, weekly, and daily time-domain characteristic analysis on the classified loads, calculate stability and predictability, and complete the classification. S8. Based on the sliding window significance test, quantify the correlation, instantaneous sensitivity and fluctuation contribution of classified load and total load; S9. Construct customized prediction models according to stability predictability classification to obtain load prediction results for each category; S10. The classification prediction results are optimally synthesized using dynamic weighting or cointegration theory, and the total load prediction value of the system is output. The photovoltaic output prediction value of S4 and the total load prediction value of S10 are then matched.

4. The distributed photovoltaic short-term forecasting method based on an artificial intelligence deep spatiotemporal model according to claim 1, characterized in that, The multi-source data in S5 is taken from photovoltaic power plants. The collected data includes daily maximum load, daily minimum load, daily peak-to-valley difference, peak-to-valley difference rate, daily average load, daily load rate, daily load fluctuation coefficient, daily peak hours, and daily median load. The collection time granularity is 15 minutes, with 96 collection points in 24 hours. Meteorological side: meteorological forecast data, including temperature, humidity, wind speed, wind direction, precipitation, and air pressure. Auxiliary information side: Structured information such as holidays, special events, planned maintenance, and load transfer records; The refined load classification in S6 includes: agricultural irrigation and drainage, power and heat production and supply, steel processing and smelting, silicon carbide, cement manufacturing, and other industries.

5. The distributed photovoltaic short-term forecasting method based on an artificial intelligence deep spatiotemporal model according to claim 4, characterized in that, The extraction of the daily load curve in S6 includes: normalizing the time-domain features of days, months, and years, using t-SNE dimensionality reduction to improve K-means++ clustering and determining the optimal number of clusters through the silhouette coefficient, and extracting typical curves according to the proportion of days.

6. The distributed photovoltaic short-term forecasting method based on an artificial intelligence deep spatiotemporal model according to claim 1, characterized in that, The S8 calculation of stability and predictability and the subsequent classification are specifically obtained through a one-sided forward difference approximation in a 15-minute discrete sampling environment under continuous-time limits. Let the sampling interval Δt = 15 minutes, then the instantaneous elasticity is εᵢ. ,t ≈(ΔP total,t / P total,t ]÷(ΔP i,t / P i,t ) Where Δ t =15 minutes, P total,t For the total load, ΔP total,t For changes in total load, P i,t For the i-th type of load, ΔP i,t Let i be the category of load change.

7. The method for improving the accuracy of total load prediction based on refined classification load characteristics according to claim 4, characterized in that, In S9, load stability is measured by the load variation coefficient, i.e. Where σ is the standard deviation of the load sequence of load i at time t, and μ is the standard deviation of the load sequence of load i at time t; the larger the value, the higher the load stability; the smaller the value, the more susceptible the load is to the influence of the load side, the meteorological side, and the auxiliary information side. The hierarchical prediction model is as follows: The PCA-SARIMAX model is used for highly stable and predictable loads; The stable and predictable load adopts the Stacking ensemble learning model; The low-stability, predictable load adopts the ConvLSTM-Transformer two-path spatiotemporal model.

8. The method for improving the accuracy of total load prediction based on refined classification load characteristics according to claim 4, characterized in that, In S10, the optimal synthesis adopts dynamic weighted ensemble learning, which uses classification load contribution and sensitivity as weights for weighted fusion.

Images