Intelligent optimization system and method for roof photovoltaic, user side energy storage and charging load coordination in clean energy-carrying park

Abstract

The invention discloses an intelligent optimization system and method for roof photovoltaic, user side energy storage and charging load coordination in a clean energy-carrying park, and belongs to the technical field of new energy and intelligent power grids. According to the system, photovoltaic power generation, an energy storage system and an intelligent charging pile are integrated, a space-time attention mechanism and a time convolutional network are adopted to predict photovoltaic output and charging requirements, and dynamic optimization scheduling is realized by combining an improved MOEA / D algorithm. An electric vehicle parking time length probability model is innovatively introduced to improve the prediction precision, meanwhile, a V2G bidirectional control strategy is designed, and SOCgt is scheduled in the peak period of electricity price; 50% of the electric vehicle reversely supplies power, and the service life of the battery is protected by adopting an SOH sensing mechanism. System hardware comprises a high-efficiency monocrystalline silicon photovoltaic module, lithium iron phosphate battery energy storage and a charging pile supporting V2G, and a software level realizes preferential supply of photovoltaic electric energy to a production load and dynamic distribution of residual electric energy to electric vehicle charging and energy storage through real-time data acquisition, prediction optimization and closed-loop control.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of new energy and smart grid, more particularly relates to a clean energy-carrying park roof photovoltaic, user-side energy storage and charging load collaborative intelligent optimization system and method. BACKGROUND

[0002] A clean energy-carrying park is a modern industrial agglomeration area with green and low carbon as the core. Its core is to rely on renewable energy (such as wind power, photovoltaic, and hydropower) or clean energy (such as nuclear power and natural gas) as the main energy supply basis, to concentrate the layout of energy-intensive industries (such as data centers, new material manufacturing, green chemical industry, and high-end equipment manufacturing), and to integrate the application of smart grid, efficient energy storage, energy cascade utilization, and digital energy efficiency management. Advanced technologies realize efficient coordination and optimal allocation of regional energy production, transmission, use, and storage (i.e., "source-grid-load-storage integration"), thereby significantly reducing unit energy consumption and carbon emission intensity. It is particularly crucial that such parks usually focus on the key link of the green power generation industry chain - the core products themselves are photovoltaic panels, wind turbine blades and towers, hydropower equipment parts, energy storage batteries, and ultra-high voltage power transmission equipment, etc. These products are directly used to build new renewable energy power generation facilities such as wind, light, and water. Therefore, the clean energy-carrying park not only highly depends on and promotes clean energy consumption, but also the "green industrial products" produced by it are indispensable components of downstream green power projects, realizing a double green closed loop from "producing with green power" to "producing green power equipment". This industrial ecosystem deeply integrated with clean energy supply, high-energy-efficiency green power equipment manufacturing, smart energy management, and low-carbon infrastructure is an important carrier for promoting green and low-carbon transformation in the industrial field and achieving the "double carbon" goal, forming a self-reinforcing green industrial closed loop.

[0003] With the rapid development of renewable energy technology, distributed photovoltaic power generation, user-side energy storage, and electric vehicle charging load have become important components of the energy internet system. Especially in clean energy-carrying parks, smart parks, and other application scenarios, the deep integration and collaborative optimization of roof photovoltaic power generation systems, battery energy storage systems, and electric vehicles and other types of loads have become a key technology direction for improving energy utilization efficiency, reducing operating costs, and promoting green and low-carbon transformation.

[0004] Existing smart energy management systems typically rely on traditional data processing and scheduling methods. Firstly, in the data processing stage, most solutions on the market employ traditional data stream processing models such as ETL (Extract-Transform-Load). While these can achieve basic energy consumption data collection and organization, they often suffer from poor compatibility and low fusion efficiency when dealing with heterogeneous data from multiple sources such as photovoltaics, batteries, and charging piles. Furthermore, their ability to automatically identify and repair abnormal data is limited, and they lack in-depth analysis of the spatiotemporal distribution characteristics of the data, affecting the accuracy of data analysis and subsequent decision-making.

[0005] Secondly, existing load and power generation forecasting models are mostly based on single variables or short-term time-series data, such as traditional methods like ARIMA and support vector machines. These methods have low prediction accuracy when dealing with long-term, multivariate, and highly dynamic energy data, and are unable to effectively capture complex situations such as the fluctuation characteristics of photovoltaic output in the park and the instantaneous surge of electric vehicle charging load, thus affecting the scheduling response time and reliability of the energy management system.

[0006] Furthermore, most current mainstream optimization and scheduling strategies employ static, single-objective, or simplified constraint mathematical models, resulting in weak adaptability to multi-source, multi-objective, and dynamically changing scenarios. For example, they lack multi-objective collaborative optimization and system self-adaptation capabilities in addressing real-time electricity price fluctuations, energy storage status adjustments, and bidirectional charging and discharging control of electric vehicles. This leads to the inability of park-level energy systems to fully respond to the external energy environment and diverse user needs in actual operation, significantly impacting the overall economic efficiency, reliability, and cleanliness of the system.

[0007] In summary, existing technologies have significant limitations in data fusion, load forecasting, and collaborative optimization, making it difficult to meet the intelligent, refined, and dynamic management needs of park-level energy systems under the background of high-proportion renewable energy access. Therefore, there is an urgent need to develop an intelligent optimization system and method that can efficiently integrate multi-source data, improve forecasting accuracy, and possess multi-objective collaborative optimization capabilities to support the efficient operation and green development of next-generation clean energy parks. Summary of the Invention

[0008] This invention proposes a systematic solution to the aforementioned technical challenges. Through an innovative data processing architecture, it achieves efficient fusion and feature enhancement of multi-source heterogeneous data, significantly improving data processing quality; it constructs a deep learning prediction model with spatiotemporal feature extraction capabilities, greatly improving the prediction accuracy of photovoltaic power output and electric vehicle charging demand; it designs a dynamic adaptive multi-objective optimization scheduling algorithm to effectively balance multiple objectives such as grid power purchase costs and battery life; and it establishes a real-time monitoring and retrospective evaluation mechanism based on digital twins, comprehensively enhancing the system's reliability and fault handling capabilities.

[0009] To achieve the above objectives, the present invention employs the following technical solution: the method comprises:

[0010] Assumptions are set for the intelligent energy management model of photovoltaic power generation system, battery energy storage system and electric vehicle charging facility;

[0011] Establish intelligent energy management models for photovoltaic power generation systems, battery energy storage systems, and electric vehicle charging facilities, including objective functions and constraints;

[0012] Data processing includes data cleaning, feature engineering, and data standardization;

[0013] The design algorithm for solving the model adopts an improved multi-objective optimization scheduling algorithm.

[0014] In one approach, the assumptions include:

[0015] (1) Time discretization: Divide a day into time intervals, total time period;

[0016] (2) Energy conservation: The system input energy is equal to the sum of the changes in output energy and stored energy;

[0017] (3) Equipment efficiency: Photovoltaic conversion efficiency, battery charging and discharging efficiency, and charging pile conversion efficiency are all time-dependent functions;

[0018] (4) Uncertainty: Photovoltaic power output and charging demand are random variables that follow a specific probability distribution.

[0019] In one scheme, the objective function includes: an economic objective function, a reliability objective function, and a comprehensive multi-objective function.

[0020] In one scheme, the constraints include: energy storage system constraints and electric vehicle charging constraints.

[0021] In one approach, the feature engineering includes: spatial feature construction mainly revolves around in-depth analysis of geographic location information, and the calculated distance features can effectively measure the degree of spatial correlation between different data collection points;

[0022] Statistical feature extraction can reveal the distribution characteristics and trends of data at a macro level, providing a statistical basis for the model.

[0023] Correlation analysis aims to uncover potential relationships between different variables, helping to screen key features and avoid feature redundancy, thereby improving model efficiency and accuracy.

[0024] In one scheme, the improved multi-objective optimization scheduling algorithm includes: the core optimization adopts the MOEA / D-DE algorithm, the differential evolution mutation operator is introduced, and the DE / best / 2 strategy is used to generate mutation vectors;

[0025] The energy storage system uses the DQN reinforcement learning algorithm to optimize the charging and discharging strategy;

[0026] The V2G bidirectional control adopts a segmented strategy, and the power is dynamically adjusted according to the photovoltaic surplus during the charging phase.

[0027] In one approach, the model employs a spatiotemporal attention-temporal convolutional network fusion model, which includes: achieving deep mining of data features through a three-layer architecture: firstly, utilizing the dilated causal convolutional properties of the temporal convolutional network to extract long-term temporal dependency features from the historical 72-hour photovoltaic power data, effectively capturing power output fluctuations caused by factors such as sudden weather changes;

[0028] Secondly, by using a bidirectional gated cyclic unit to synchronously process forward and reverse timing information, the identification of periodic load patterns such as morning and evening peak hours is enhanced.

[0029] Ultimately, a spatiotemporal attention mechanism is used to assign dynamic weights to key influencing factors such as solar irradiance, temperature, and weekdays / holidays.

[0030] To predict electric vehicle charging demand, a Weibull distribution probability model is introduced to characterize vehicle parking time, and a hybrid prediction model is constructed by combining user travel behavior data.

[0031] In one approach, the spatiotemporal attention-temporal convolutional network fusion model integrates the photovoltaic power output and charging demand prediction, and the temporal convolutional network (TCN) processes historical data through dilated causal convolution.

[0032] To address the charging needs of electric vehicles, a Weibull distribution probability model is constructed to describe the parking duration.

[0033] On another front, a smart optimization system for the coordinated operation of rooftop photovoltaics, user-side energy storage, and charging load in a clean energy-intensive park is provided. The system is applicable to the method described above. The system includes: an observable detection module for collecting data on electricity price fluctuations and the energy storage status of the battery energy storage system, wherein the battery energy storage system collects electricity consumption data of electric vehicle charging facilities.

[0034] The clean energy park power collaborative dispatch system is used to collect the power consumption demand of production power-consuming units, the power generation data of photovoltaic power generation units, and to transmit and interact with the power grid on electricity price information.

[0035] The virtual power plant module interacts with the clean energy park's power dispatch system for command exchange and feedback adjustment.

[0036] Beneficial effects of this invention:

[0037] This system innovatively introduces a probabilistic model of electric vehicle parking duration (Weibull distribution) to improve prediction accuracy. It also designs a V2G bidirectional control strategy to schedule electric vehicles with a State of Charge (SOC) > 50% to supply power during peak electricity price periods, and employs a State of Health (SOH) sensing mechanism to protect battery life. The system hardware includes high-efficiency monocrystalline silicon photovoltaic modules, lithium iron phosphate battery energy storage, and V2G-enabled charging piles. At the software level, real-time data acquisition, predictive optimization, and closed-loop control prioritize photovoltaic power supply to production loads, with surplus power dynamically allocated to electric vehicle charging and energy storage.

[0038] This invention proposes a systematic solution to the aforementioned technical challenges. Through an innovative data processing architecture, it achieves efficient fusion and feature enhancement of multi-source heterogeneous data, significantly improving data processing quality; it constructs a deep learning prediction model with spatiotemporal feature extraction capabilities, greatly improving the prediction accuracy of photovoltaic power output and electric vehicle charging demand; it designs a dynamic adaptive multi-objective optimization scheduling algorithm, effectively balancing multiple objectives such as grid power purchase costs and battery life; and it establishes a real-time monitoring and retrospective evaluation mechanism based on digital twins, comprehensively enhancing the system's reliability and fault handling capabilities. These technologies are closely interconnected. Deficiencies in data processing directly affect the input quality of the prediction model, leading to inaccurate prediction results. Inaccurate predictions, in turn, lack a reliable basis for optimal scheduling, further reducing the effectiveness of scheduling strategies. Simultaneously, deficiencies in scheduling strategies, in turn, amplify errors in data processing and prediction, creating a vicious cycle. This mutually restrictive relationship makes it difficult for traditional energy management systems to operate efficiently.

[0039] This invention comprehensively optimizes the energy management system by constructing a complete technical chain, from data processing, predictive analysis, and optimized scheduling to system monitoring. Through the synergistic effect of each link, it breaks through traditional technical bottlenecks, achieves intelligent upgrades in energy management, significantly improves the overall performance and operational efficiency of the system, and provides strong support for the efficient utilization of clean energy and the stable operation of the power grid. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the model construction process of the present invention.

[0041] Figure 2 This is a diagram of the improved algorithm framework of the present invention;

[0042] Figure 3 This is a schematic diagram of the power coordinated dispatch system architecture of the present invention;

[0043] Figure 4 This is a schematic diagram of the multi-objective optimization results of the present invention. Detailed Implementation

[0044] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0045] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0046] In the field of intelligent energy management system technology, traditional data processing, prediction models, and optimization scheduling methods have significant limitations. Existing ETL data processing methods struggle to efficiently integrate heterogeneous data from multiple sources such as photovoltaics, energy storage, and electric vehicles, and are insufficient in anomaly repair and spatiotemporal feature extraction. Prediction models often rely on single variables, resulting in low accuracy in processing long-sequence data and failing to accurately capture complex energy change patterns. Optimization scheduling strategies lack multi-objective coordination and dynamic adaptability, making it difficult to cope with real-time changes in electricity prices and user demand. These intertwined problems form a technological bottleneck, hindering the intelligent development of energy management systems.

[0047] This invention proposes a systematic solution to the aforementioned technical challenges. Through an innovative data processing architecture, it achieves efficient fusion and feature enhancement of multi-source heterogeneous data, significantly improving data processing quality; it constructs a deep learning prediction model with spatiotemporal feature extraction capabilities, greatly improving the prediction accuracy of photovoltaic power output and electric vehicle charging demand; it designs a dynamic adaptive multi-objective optimization scheduling algorithm, effectively balancing multiple objectives such as grid power purchase costs and battery life; and it establishes a real-time monitoring and retrospective evaluation mechanism based on digital twins, comprehensively enhancing the system's reliability and fault handling capabilities. These technologies are closely interconnected. Deficiencies in data processing directly affect the input quality of the prediction model, leading to inaccurate prediction results. Inaccurate predictions, in turn, lack a reliable basis for optimal scheduling, further reducing the effectiveness of scheduling strategies. Simultaneously, deficiencies in scheduling strategies, in turn, amplify errors in data processing and prediction, creating a vicious cycle. This mutually restrictive relationship makes it difficult for traditional energy management systems to operate efficiently.

[0048] A smart optimization method for coordinating rooftop photovoltaics, user-side energy storage, and charging loads in clean energy-intensive industrial parks, such as... Figure 1 As shown, step 1 sets the assumptions for the intelligent energy management model of the photovoltaic power generation system, battery energy storage system, and electric vehicle charging facilities;

[0049] This model is for a smart energy management system that includes a photovoltaic power generation system, a battery energy storage system (BESS), and electric vehicle charging facilities, and considers the following assumptions:

[0050] (1) Time discretization: Divide a day into time intervals (e.g., 15 minutes) and total time periods.

[0051] (2) Energy conservation: The system input energy is equal to the sum of the changes in output energy and stored energy.

[0052] (3) Equipment efficiency: Photovoltaic conversion efficiency, battery charging and discharging efficiency, and charging pile conversion efficiency are all time-dependent functions.

[0053] (4) Uncertainty: Photovoltaic power output and charging demand are random variables that follow a specific probability distribution.

[0054] Step 2: Establish an intelligent energy management model for the photovoltaic power generation system, battery energy storage system, and electric vehicle charging facilities, including the objective function and constraints;

[0055] (1) Economic objectives:

[0056] Minimize the total system cost C total ,include

[0057] Electricity purchase cost from the grid:

[0058]

[0059] Battery degradation cost:

[0060]

[0061] Equipment maintenance costs:

[0062]

[0063] Where λ t Let P be the electricity price during time period t. grid (t) represents the power purchased from the grid. N represents the energy storage charging / discharging power, and N represents the number of charging stations.

[0064] (2) Reliability objectives:

[0065] Maximize system reliability R system Measured by the following indicators:

[0066] Power supply reliability:

[0067]

[0068] Where U(t) represents the unmet demand in time period t, and D(t) represents the total demand.

[0069] Equipment availability:

[0070]

[0071] (3) Integrating multi-objective functions

[0072] Using the weighted summation method:

[0073]

[0074] Where w1 + w2 = 1, This is the normalized baseline value.

[0075] (4) Constraints:

[0076]

[0077] Among them, P loss (t) represents the system power loss, including line loss and conversion loss.

[0078] Constraints of energy storage systems:

[0079] SOC dynamic equation:

[0080]

[0081] SOC boundary constraints:

[0082] SOC min ≤SOC ES (t)≤SOC max (9)

[0083] Charge and discharge power limits:

[0084]

[0085] Constraints of photovoltaic power generation:

[0086] Output Limitation:

[0087]

[0088] Prediction error constraints:

[0089]

[0090] Where, ε PV(t) represents the prediction error threshold.

[0091] Electric vehicle charging constraints:

[0092] Charging needs met:

[0093]

[0094] Charging station power limitations:

[0095]

[0096] V2G constraint: If SOC EV (t)≥SOC V2G ,but Power grid interaction constraints:

[0097] Maximum power purchase capacity:

[0098]

[0099] Peak-valley time constraints: λ t ∈{λ p ,λ v ,λ n It is determined based on the time period t.

[0100] (5) Key model construction and uncertainty handling

[0101] The TCN-BiGRU-Attention model is adopted:

[0102]

[0103] Where G(t) is solar irradiance, T(t) is temperature, Time(t) is time characteristic, and Weather(t) is meteorological conditions (such as humidity and cloud cover).

[0104] Electric vehicle charging demand prediction model, based on user behavior analysis and probability distribution:

[0105]

[0106] Where p(D|t,DayType,Location) is the probability density function of charging demand given time, date type, and location.

[0107] The battery degradation model uses the rainflow counting method combined with the Arrhenius equation:

[0108]

[0109] Among them, C i Depth of charge / discharge cycles To correspond to the cycle life, E a The activation energy is T, where k is the Boltzmann constant. ref This is a reference temperature.

[0110] For handling uncertainty, a stochastic programming method is used, treating photovoltaic power output and charging demand as random variables:

[0111] min[F]=min∫ Ω F(ω)p(ω)dω(19)

[0112] Where Ω is the set of all possible scenarios, and p(ω) is the probability distribution of scenario ω.

[0113] Step 3: Data processing, including data cleaning, feature engineering, and data standardization;

[0114] (1) Data cleaning

[0115] In the data preprocessing stage of solving intelligent energy management system models, data cleaning is a core step in ensuring data quality. Its purpose is to eliminate noise, missing values, and outliers in multi-source heterogeneous data, providing a reliable data foundation for subsequent model training and decision-making. Because data sources such as photovoltaic power, energy storage SOC, and electric vehicle charging demand are widespread and susceptible to factors such as sensor failures, communication interruptions, and environmental interference, data cleaning is a complex and critical task.

[0116] In handling missing values, traditional simple methods such as mean imputation and median imputation can damage the original distribution characteristics and temporal correlation of the data, making them unsuitable for complex models. This system employs Spatiotemporal Weighted Regression (STWR) for missing value repair. This method integrates regression analysis with a spatiotemporal weighting mechanism, fully considering the temporal correlation and spatial proximity of the data. Its core idea is to utilize the spatiotemporal neighborhood information surrounding the target data point, combined with external auxiliary variables related to the target data (such as the impact of meteorological data on photovoltaic output), to perform weighted regression prediction. The mathematical expression is as follows:

[0117]

[0118] Among them For missing value estimation, w k It is a Gaussian function The calculated spatiotemporal weights are based on the spatial distance d. s and time interval d t Dynamic adjustments are made to ensure that nearest neighbors contribute more to the prediction; x k,j For neighborhood data points, z i,l As an auxiliary variable, β l Then, the least squares method is used. In its implementation, the method first constructs a spatiotemporal neighborhood, defines a spatiotemporal search radius to select neighborhood data, then calculates weights, estimates regression parameters, and finally completes the missing value prediction. In practical applications, compared to traditional methods, this approach significantly reduces the root mean square error (RMSE) of missing value repair in the processing of measured data from a photovoltaic power station, thus significantly improving data integrity.

[0119] Outlier detection is crucial for ensuring model reliability, as sudden false alarms from sensors and abrupt changes in grid voltage can severely interfere with model training, leading to the failure of energy dispatch strategies. This system employs a two-layer detection framework combining Isolation Forest and the 3σ criterion. Isolation Forest, an unsupervised learning algorithm, is based on the principle that "outliers are more easily isolated." It constructs a binary tree by randomly sampling samples from the original data to isolate data points. During tree construction, outliers, due to their uniqueness, are often isolated to leaf nodes with fewer splits. Anomaly scores are obtained by calculating the path length of each data point. However, Isolation Forest may misclassify data when dealing with ambiguous boundaries and complex data distributions. Therefore, the 3σ criterion is introduced for secondary verification. The 3σ criterion is based on the assumption of a normal distribution, assuming that data falls within the mean. The probability of values ​​outside the range of σ (plus or minus three standard deviations) is extremely small; data outside this range are considered outliers. Combining these two methods leverages the efficiency of Isolation Forest to quickly identify potential anomalies while improving detection accuracy through the 3σ criterion, effectively avoiding the limitations of a single method.

[0120] Data smoothing aims to eliminate random noise in data, restore the true trend of data changes, and ensure the continuity and stability of the data. This system uses adaptive Kalman filtering to achieve data smoothing. This method is based on a state-space model and continuously optimizes the data estimate through two steps: prediction and update. In the prediction phase, based on the state estimate x from the previous time step... k-1|k-1 and the system state transition matrix F k Combined with process noise w k and control input u k The predicted state value at the current moment is obtained.

[0121] x k|k-1 =F k x k-1|k-1 +B k u k +w k (twenty one)

[0122] At the same time, it is also necessary to update the covariance of the prediction error in a timely manner. During the update phase, the observed value z is used. k and observation matrix H k Through Kalman gain

[0123]

[0124] The predicted values ​​are corrected to obtain a more accurate state estimate.

[0125] x k|k =x k|k-1 +K k (z k -H k x k|k-1 )(twenty three)

[0126] And update the error covariance in a timely manner.

[0127] p k|k =(IK k H k )p k|k-1 (twenty four)

[0128] Compared to traditional smoothing methods such as moving average, adaptive Kalman filtering can adjust filtering parameters in real time according to the dynamic changes of data, resulting in better noise suppression. It is particularly suitable for data scenarios with frequent fluctuations and high uncertainty in intelligent energy management systems, effectively improving data availability and model stability. Through comprehensive processing including missing value handling, outlier detection, and data smoothing, the system's data cleaning process comprehensively improves data quality, laying a solid foundation for accurate model solving and efficient operation of the energy system.

[0129] (2) Feature Engineering

[0130] In the model solving process of intelligent energy management systems, feature engineering serves as a crucial bridge connecting raw data and model applications. It plays a vital role in transforming diverse and heterogeneous raw data (covering photovoltaic power, energy storage system status, electric vehicle charging demand, etc.) into high-quality features that the model can effectively utilize. Raw data often suffers from complexity, information redundancy, and obscure correlations, making it difficult to directly adapt to model requirements. Therefore, in-depth feature engineering work is needed from multiple dimensions, including spatiotemporal characteristic mining, statistical regularity extraction, and variable relationship analysis. This enhances the data's ability to express the operational laws of the energy system, thereby improving the accuracy and reliability of model prediction and optimized scheduling.

[0131] In terms of constructing time features, meticulously breaking down timestamps is a key element. Precisely dividing a day into 24-hour increments clearly reveals the cyclical patterns of data changes throughout the day. For example, the trend of photovoltaic power increasing during the day with stronger sunlight and gradually decreasing at night can be intuitively presented through this feature. Considering the different characteristics of the seven days of the week effectively distinguishes between weekdays and rest days. Since people's travel habits vary significantly across different dates, electric vehicle charging demand exhibits a clear cyclical change, and this feature helps the model accurately identify such patterns. Dividing a year into four seasons, the duration, intensity, and temperature of sunlight in different seasons not only significantly impact photovoltaic power generation efficiency and energy storage device performance but also change people's travel patterns, ultimately affecting electric vehicle charging demand. This feature provides strong support for the model to consider the comprehensive impact of seasonal factors on the energy system.

[0132] Spatial feature construction primarily revolves around in-depth analysis of geographic location information. This is achieved through formulas...

[0133]

[0134] The calculated distance features can effectively measure the spatial correlation between different data collection points. In distributed photovoltaic (PV) power generation scenarios, adjacent PV power plants often exhibit a strong correlation in power generation due to similar meteorological conditions; distance features can quantify this spatial correlation.

[0135]

[0136] The defined directional features further clarify the locational relationships between data collection points. For example, photovoltaic panels in different locations have significantly different power generation capacities due to differences in the angle at which they receive solar radiation. Directional features help the model fully consider the impact of this spatial location factor on power generation. Furthermore, combining spatial features such as altitude and terrain slope from Geographic Information System (GIS) data further refines the dimensions of spatial factors' influence on energy data. Cross-combining temporal and spatial features, such as combinations like "season-distance" and "hours per day-direction," allows for a more comprehensive depiction of energy data variation patterns from a spatiotemporal perspective, providing the model with richer and more discriminative input information.

[0137] Statistical feature extraction can reveal the distribution characteristics and trends of data at a macro level, providing a statistical basis for models. Sliding window statistics are commonly used and effective methods, among which the mean...

[0138]

[0139] It reflects the average level of data within a sliding window of length k. For example, by calculating the average photovoltaic power over the past hour, the power generation trend in the near future can be predicted; standard deviation

[0140]

[0141] This measures the dispersion of data within a window. A large standard deviation of photovoltaic power within a certain period indicates that the power generation fluctuates drastically during that period, possibly due to factors such as sudden weather changes; trend.

[0142]

[0143] This describes the trend of data changes within the window, which is of significant reference value for predicting future trends in energy data. By adjusting the sliding window size k, statistical characteristics of the data at different time scales can be obtained, meeting the model's need for multi-scale information. Wavelet transform feature extraction.

[0144]

[0145] This method can decompose raw data into components of different frequencies. High frequencies correspond to instantaneous fluctuations in the data, while low frequencies reflect long-term trends. This allows the model to better capture the patterns of data change across different time scales and frequencies, making it particularly suitable for processing non-stationary energy data. Furthermore, higher-order statistics such as skewness and kurtosis can further characterize the data distribution, providing the model with more detailed information. For example, if the kurtosis of the energy storage device's charging and discharging power data is high during a certain period, it indicates the presence of many extreme values, possibly due to sudden changes in grid load or equipment failure. The model can then make more informed decisions based on this information.

[0146] Correlation analysis aims to uncover potential relationships between different variables, helping to screen key features and avoid feature redundancy, thereby improving model efficiency and accuracy. Pearson correlation coefficient.

[0147]

[0148] By measuring the linear correlation between variables, a correlation coefficient matrix can be constructed to show the degree of association between the variables. Based on this, a feature association network is built, and thresholds are set to determine the connectivity between features.

[0149]

[0150] It presents complex relationships between features graphically, facilitating the identification of key features and the handling of redundant features. Furthermore, it includes interactive information.

[0151]

[0152] Nonlinear measurement methods can uncover nonlinear relationships between variables, further enriching the information sources for feature engineering.

[0153] (3) Data standardization

[0154] In the data preprocessing process for solving intelligent energy management system models, data standardization is a crucial step in ensuring data quality and improving model training effectiveness. Raw, multi-source, heterogeneous data (such as photovoltaic power, energy storage SOC, and electric vehicle charging demand) have different dimensions and large differences in numerical distribution. Directly inputting these into the model can easily lead to slow training convergence, parameter imbalance, and even model failure. Therefore, standardization is necessary to unify the data scale and enhance the comparability of data features and the adaptability of the model. Traditional standardization methods (such as min-max standardization and Z-score standardization) have limitations when processing energy data. Because energy data has significant spatiotemporal characteristics and non-stationary fluctuations, simple global standardization may mask local data patterns and weaken the effective information of the data. This model adopts a "spatiotemporal weighted standardization" method, fully considering the temporal correlation and spatial proximity of the data, and achieving refined data normalization through dynamic weight allocation. The core of spatiotemporal weighted standardization lies in assigning different weights to data points based on their spatiotemporal location for standardization calculation. In the temporal dimension, data from nearby moments tend to have higher correlations. For example, photovoltaic power changes in adjacent time periods on the same day show similar trends, so recent data is given higher weight. In the spatial dimension, data collected by sensors in geographically close locations are affected by the same environmental factors (such as light intensity and temperature in the same area), and data points that are closer together have higher weights in the standardization process. In this way, the standardized features can retain the spatiotemporal characteristics of the original data while reflecting the relative changes of the data at a uniform scale. In the specific implementation process, the spatiotemporal search radius is first defined to determine the spatiotemporal neighborhood of each data point. Within this neighborhood, the corresponding weights are calculated based on the time interval and spatial distance between the data point and the target point. The shorter the time interval and the closer the spatial distance, the greater the weight of the data point. Then, the local mean and standard deviation are calculated based on these weights, which serve as the benchmark for standardization. Compared with traditional methods that use global statistics for standardization, spatiotemporal weighted standardization can better adapt to the fluctuation characteristics of energy data in different time periods and regions, avoiding feature distortion caused by global standardization. In addition, to cope with abnormal fluctuations and noise interference in energy data, the standardization process also introduces the concept of robust statistics. When calculating the mean and standard deviation, robust estimation methods are employed to reduce the impact of outliers on the standardized parameters. For example, the median is used instead of the mean, and the M-estimator is used to calculate the standard deviation, ensuring the stability and reliability of the standardization process. This approach ensures that the standardized data reflects the true data distribution characteristics while reducing the negative impact of noise on model training. Through spatiotemporal weighted standardization, energy data of different types and magnitudes are unified into similar numerical ranges, eliminating interference caused by differences in units. Simultaneously, dynamic weighting based on the spatiotemporal characteristics of the data effectively preserves the inherent patterns and trends of the data, providing high-quality input data for subsequent prediction models and optimization algorithms.Practical applications show that adopting this standardization method improves the convergence speed of model training and the prediction accuracy, significantly enhancing the performance and reliability of the intelligent energy management system.

[0155] Step 4: Design the model solution algorithm, using an improved multi-objective optimization scheduling algorithm. For example... Figure 2 As shown.

[0156] (1) Prediction algorithm system based on multi-source data fusion

[0157] The system employs a spatiotemporal attention-temporal convolutional network fusion model to achieve accurate prediction of photovoltaic power output and charging demand. This algorithm overcomes the limitations of traditional single-model approaches by using a three-layer architecture to deeply mine data features: First, it leverages the dilated causal convolution characteristics of the temporal convolutional network (TCN) to extract long-term temporal dependency features from historical 72-hour photovoltaic power data, effectively capturing power output fluctuations caused by factors such as sudden weather changes; second, it uses a bidirectional gated recurrent unit (BiGRU) to simultaneously process forward and reverse temporal information, enhancing the identification of periodic load patterns such as morning and evening peak hours; finally, through a spatiotemporal attention mechanism, it assigns dynamic weights to key influencing factors such as solar irradiance, temperature, and weekdays / holidays, automatically increasing the weight of meteorological data during cloudy weather. To predict electric vehicle charging demand, an innovative Weibull distribution probability model was introduced to characterize vehicle parking duration. By fitting historical parking data of 2,000 electric vehicles in a certain region, the optimal distribution parameters with shape parameter k=2.3 and scale parameter λ=3.8 were determined. Combined with user travel behavior data (such as weekday commuting patterns), a hybrid prediction model was constructed, which controlled the charging demand prediction error within 4.1%, improving the accuracy by 37% compared with the traditional ARIMA model.

[0158] The system employs a TCN-BiGRU-Attention ensemble model to predict photovoltaic power output and charging demand. The Temporal Convolutional Network (TCN) processes historical data through dilated causal convolutions; its core computation is as follows:

[0159]

[0160] Where W is the convolution kernel weight, b is the bias term, and σ is the ReLU activation function, the receptive field is expanded by the dilation factor to capture long-term temporal dependencies. The Bidirectional Gated Recurrent Unit (BiGRU) processes both forward and backward temporal features simultaneously, and the forward hidden state update formula is:

[0161]

[0162] The backward hidden state is computed with input in reverse order, and the final output is a concatenation of bidirectional features. The spatiotemporal attention mechanism computes weights using a Query-Key-Value structure.

[0163]

[0164] Among them W q and W k Let h be the projection matrix. t As a time-series feature vector, this mechanism enables the model to focus on key influencing factors (such as meteorological data during extreme weather periods).

[0165] To address the charging needs of electric vehicles, a Weibull distribution probability model is constructed to describe the parking duration T. p :

[0166]

[0167] The parameters λ (scale parameter) and k (shape parameter) are determined by fitting historical data using maximum likelihood estimation, and then combined with user behavior features B. t (e.g., weekday commuting mode) and historical charging data C hist The final prediction model is:

[0168]

[0169] (2) Improved multi-objective optimization scheduling algorithm

[0170] like Figure 4 The core optimization module adopts the MOEA / D-DE algorithm, with the grid purchase cost F1 and battery degradation cost F2 as optimization objectives.

[0171] minF=ω1F1+ω2F2 (39)

[0172] in:

[0173]

[0174] c t For time-of-use electricity pricing, P grid,t Where η is the power purchased, η is the attenuation coefficient, and ΔSOC is the power consumption. t For the change in SOC of energy storage, SOH t This represents the current battery health status, while SOH0 represents the initial health status.

[0175] The algorithm introduces a differential evolution mutation operator and uses a DE / best / 2 strategy to generate mutation vectors:

[0176] v i,g =x best,g +F·(x r1,g -x r2,g )+F·(x r3,g -x r4,g (41)

[0177] Where xbest,g For the current optimal solution, x r1,g ,x r2,g ,x r3,g ,x r4,g For random individuals, F is a scaling factor (usually 0.8). The dynamic weight adjustment mechanism updates in real time based on real-time electricity price fluctuations.

[0178]

[0179] in γ is the initial weight (default 0.6), γ is the adjustment coefficient (0.5), and c is the average daily electricity price.

[0180] The constraints include:

[0181] Energy storage SOC boundary:

[0182] SOC min ≤SOC t ≤SOC max (Default 20%-80%) (43)

[0183] Charging station power limitations:

[0184] 0≤P ch,t ≤P ch,max (44)

[0185] V2G discharge conditions:

[0186] SOC t >50%∧SOH t >80% (45)

[0187] (3) Energy storage and V2G coordinated control algorithm

[0188] The energy storage system employs the DQN reinforcement learning algorithm to optimize the charging and discharging strategy, and the state space is defined as follows:

[0189] S t =[SOC t ,P PV,t D t ,c t (46)

[0190] Among them, D t This represents the load demand, while the operating range refers to the charging / discharging power level (A). t ∈{-P max ,...,0,...,P max The reward function is designed as follows:

[0191]

[0192] in For loads not met by photovoltaic (PV) grid connection, α is the PV grid connection bonus coefficient (0.3), and β is the SOC fluctuation penalty coefficient (0.1). The Q network updates parameters through empirical replay, and the target value is calculated as follows:

[0193] Y t =R t +γ·max A′ Q(S t+1 ,A′;θ′) (48)

[0194] Where γ is the discount factor (0.95) and θ′ is the target network parameter.

[0195] V2G bidirectional control adopts a segmented strategy, dynamically adjusting the power based on photovoltaic surplus during the charging phase:

[0196]

[0197] Where η ch Let I be the charging efficiency (0.92) and I be the indicator function. During the discharge phase, a SOH-SOC decision matrix is ​​established, where the electricity price is higher than the peak-valley threshold:

[0198]

[0199] Where k is the discharge coefficient (0.8), P v2g,max This represents the maximum discharge power.

[0200] (4) Data preprocessing algorithm principle

[0201] When using spatiotemporal weighted regression to repair missing values, for the target point Constructing a spatiotemporal neighborhood:

[0202]

[0203] Among them, weight d t d represents the time interval (default 1 hour). s This refers to spatial distance.

[0204] Anomaly detection employs a concatenation of isolated forest and the 3σ criterion: the isolated forest calculates anomaly scores by measuring path length h(x).

[0205]

[0206] An anomaly is identified when s(x) > 0.8 and the data exceeds [μ-3σ, μ+3σ]. The adaptive Kalman filter update equation is:

[0207]

[0208] Where Kt For Kalman gain, P t|t-1 This is the prediction error covariance.

[0209] Table 1 Symbols and Explanations

[0210]

[0211]

[0212]

[0213] Table 2: Advantages Analysis

[0214]

[0215] like Figure 3 As shown, an intelligent optimization system for the coordinated operation of rooftop photovoltaics, user-side energy storage, and charging loads in a clean energy industrial park is disclosed. The system includes: an observable detection module for collecting data on electricity price fluctuations and the energy storage status of the battery energy storage system, which collects electricity consumption data from electric vehicle charging facilities; a clean energy industrial park power coordination and dispatch system for collecting electricity consumption demands from production power-consuming units and power generation data from photovoltaic power generation units, and transmitting and interacting with the power grid on electricity price information; and a virtual power plant module for command interaction and feedback adjustment with the clean energy industrial park power coordination and dispatch system.

[0216] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0217] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart optimization method for the coordinated operation of rooftop photovoltaics, user-side energy storage, and charging loads in a clean energy-intensive industrial park, characterized in that: The method includes: Assumptions are set for the intelligent energy management model of photovoltaic power generation system, battery energy storage system and electric vehicle charging facility; Establish intelligent energy management models for photovoltaic power generation systems, battery energy storage systems, and electric vehicle charging facilities, including objective functions and constraints; Data processing includes data cleaning, feature engineering, and data standardization; The design algorithm for solving the model adopts an improved multi-objective optimization scheduling algorithm.

2. The intelligent optimization method for coordinating rooftop photovoltaic, user-side energy storage, and charging load in a clean energy-intensive park according to claim 1, characterized in that: The assumptions mentioned include: (1) Time discretization: Divide a day into time intervals, total time period; (2) Energy conservation: The system's input energy equals the sum of the changes in output energy and stored energy; (3) Equipment efficiency: Photovoltaic conversion efficiency, battery charging and discharging efficiency, and charging pile conversion efficiency are all time-dependent functions; (4) Uncertainty: Photovoltaic power output and charging demand are random variables that follow a specific probability distribution.

3. The intelligent optimization method for coordinating rooftop photovoltaic, user-side energy storage, and charging load in a clean energy-intensive park according to claim 1, characterized in that: The objective functions include: economic objective function, reliability objective function, and comprehensive multi-objective function.

4. The intelligent optimization method for coordinating rooftop photovoltaic, user-side energy storage, and charging load in a clean energy-intensive park according to claim 1, characterized in that: The constraints include: energy storage system constraints and electric vehicle charging constraints.

5. The intelligent optimization method for coordinating rooftop photovoltaic, user-side energy storage, and charging load in a clean energy-intensive park according to claim 1, characterized in that: The feature engineering includes: spatial feature construction mainly focuses on in-depth analysis of geographic location information, and the calculated distance features can effectively measure the degree of spatial correlation between different data collection points; Statistical feature extraction can reveal the distribution characteristics and trends of data at a macro level, providing a statistical basis for the model. Correlation analysis aims to uncover potential relationships between different variables, helping to screen key features and avoid feature redundancy, thereby improving model efficiency and accuracy.

6. The intelligent optimization method for coordinating rooftop photovoltaic, user-side energy storage, and charging load in a clean energy-intensive park according to claim 1, characterized in that: The improved multi-objective optimization scheduling algorithm includes: the core optimization adopts the MOEA / D-DE algorithm, introduces the differential evolution mutation operator, and uses the DE / best / 2 strategy to generate mutation vectors; The energy storage system employs the DQN reinforcement learning algorithm to optimize the charging and discharging strategy; The V2G bidirectional control adopts a segmented strategy, and the power is dynamically adjusted according to the photovoltaic surplus during the charging phase.

7. The intelligent optimization method for coordinating rooftop photovoltaic, user-side energy storage, and charging load in a clean energy-intensive park according to claim 1, characterized in that: The model described above employs a spatiotemporal attention-temporal convolutional network fusion model, which includes: achieving in-depth mining of data features through a three-layer architecture: firstly, utilizing the dilated causal convolutional characteristics of the temporal convolutional network to extract long-term time-series dependent features from the historical 72-hour photovoltaic power data, effectively capturing power output fluctuations caused by factors such as sudden weather changes; Secondly, by using a bidirectional gated cyclic unit to synchronously process forward and reverse timing information, the identification of periodic load patterns such as morning and evening peak hours is enhanced. Ultimately, a spatiotemporal attention mechanism is used to assign dynamic weights to key influencing factors such as solar irradiance, temperature, and weekdays / holidays. To predict electric vehicle charging demand, a Weibull distribution probability model is introduced to characterize vehicle parking time, and a hybrid prediction model is constructed by combining user travel behavior data.

8. The intelligent optimization method for coordinating rooftop photovoltaic, user-side energy storage, and charging load in a clean energy-intensive park according to claim 7, characterized in that: The spatiotemporal attention-temporal convolutional network fusion model integrates the prediction of photovoltaic power output and charging demand. The temporal convolutional network (TCN) processes historical data through dilated causal convolution. To address the charging needs of electric vehicles, a Weibull distribution probability model is constructed to describe the parking duration.

9. A smart optimization system for the coordinated operation of rooftop photovoltaics, user-side energy storage, and charging loads in a clean energy-intensive industrial park, wherein the system is applicable to the method described in any one of claims 1-8, characterized in that: The system includes: an observable detection module for collecting data on electricity price fluctuations and the energy storage status of the battery energy storage system; and the battery energy storage system for collecting electricity consumption data of electric vehicle charging facilities. The clean energy park power collaborative dispatch system is used to collect the power consumption demand of production power-consuming units, the power generation data of photovoltaic power generation units, and to transmit and interact with the power grid on electricity price information. The virtual power plant module interacts with the clean energy park's power dispatch system for command exchange and feedback adjustment.

Images