Light storage energy scheduling optimization recommendation method and system based on artificial intelligence

Abstract

The invention discloses a light storage energy scheduling optimization recommendation method based on artificial intelligence, and relates to the technical field of energy management. According to the method, power generation side state parameters, energy storage side working data, load side demand characteristics and power grid side economic signals are collected in real time, and a comprehensive energy data set is formed through space-time alignment and standardized fusion; analyzing the data set by adopting a time-space two-dimensional prediction model, and outputting the power generation capacity, the power generation fluctuation ratio, the load and the risk probability distribution in a future time period; the predicted value is input into an intelligent decision engine, and an optimal scheduling strategy is deduced and generated in a virtual simulation environment through a strategy optimization algorithm; the strategy is converted into a control instruction through a northbound interface to be issued and executed, and execution deviation is collected to generate a verification report; strategy effect data is fed back to the cloud incremental learning platform, model parameters are updated online through distributed collaborative learning, and continuous evolution and scene self-adaptive optimization of the scheduling strategy are achieved. The method has the effect of improving the economic benefit of light storage energy scheduling.

Description

Technical Field

[0001] This invention relates to the technical field of photovoltaic and energy storage management, and in particular to a method and system for optimizing and recommending photovoltaic and energy storage scheduling based on artificial intelligence. Background Technology

[0002] In the field of photovoltaic and energy storage dispatch, the coordinated operation of photovoltaic power generation and energy storage systems is the core model for achieving efficient utilization of clean energy. The basic technical framework of this field involves monitoring photovoltaic output, energy storage status, load demand, and grid signals to formulate charging and discharging plans to balance the spatiotemporal mismatch between power generation and consumption. The goal is to maximize economic benefits and equipment lifespan while ensuring power supply reliability. Current mainstream technologies generally employ rule-based control based on fixed thresholds combined with short-term prediction algorithms, using a layered architecture of Battery Management System (BMS) and Energy Management System (EMS) to issue dispatch commands.

[0003] Existing scheduling systems rely heavily on manual experience to set static parameters. Strategy formulation is limited to a single economic goal while neglecting equipment safety and lifespan loss. Furthermore, there is a lack of effective verification methods before strategy execution, which reduces the economic benefits of photovoltaic and energy storage scheduling. Therefore, improvements are needed. Summary of the Invention

[0004] To improve the economic efficiency of photovoltaic and energy storage dispatch, this application provides a method and system for optimizing photovoltaic and energy storage dispatch based on artificial intelligence.

[0005] Firstly, the above-mentioned inventive objective of this application is achieved through the following technical solution: Multimodal data from the power generation side is collected in real time, and the multimodal data is spatiotemporally aligned and standardized and fused to form a comprehensive energy dataset. The multimodal data includes state parameters, energy storage side operating data, load side demand characteristics, and grid side economic signals. The pre-set time-space dual-dimensional prediction model analyzes the comprehensive energy dataset based on a machine self-learning algorithm to output predicted power generation operating conditions data for future periods, wherein the predicted power generation operating conditions data includes predicted power generation, predicted power generation volatility, predicted load, and load risk probability distribution. The predicted values ​​are input into the intelligent decision engine, and the strategy optimization algorithm is used to perform strategy deduction in a virtual simulation environment to generate the optimal scheduling policy, wherein the optimal scheduling policy includes energy storage scheduling curve, power generation control value, and load scheduling instruction. The optimal scheduling strategy is converted into a set of control instructions that conforms to the industry communication protocol through the northbound interface and sent to the power generation equipment, energy storage system and load management terminal for execution. During the execution process, the strategy execution deviation value is collected in real time, and the residual value is calculated by comparing the deviation value with the predicted value to generate a strategy effect verification report. The strategy effectiveness verification report, actual execution data, and user feedback information are sent to the cloud-based incremental learning platform. The platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, updates the parameters of the prediction model and decision engine online, and realizes continuous evolution and scenario adaptive optimization of the scheduling strategy.

[0006] By adopting the above technical solutions, a five-layer closed loop of "multi-source perception, multi-dimensional prediction, intelligent decision-making, reliable execution, and continuous evolution" is constructed, achieving a systematic leap in scheduling strategies from passive response to proactive prediction, from static fixation to dynamic adaptation, and from single-objective optimization to multi-objective collaboration. Specifically, multi-source data fusion and spatiotemporal alignment technologies significantly improve the accuracy of power generation and load forecasting, avoiding the risks of overcharging and discharging of energy storage or load outages caused by prediction errors in traditional methods; the intelligent decision engine pre-simulates strategies in virtual simulations, ensuring that they are verified for both safety and economy before being deployed, preventing the direct execution of high-risk strategies; the cloud-based continuous learning mechanism enables the model to evolve itself with the accumulation of data from multiple systems, adapting to changes in different scenarios and avoiding the rigidity and failure of traditional models; the integrated hardware and software controller achieves a millisecond-level closed loop from perception to execution, supporting rapid response to second-level load impacts. This solution elevates the scheduling of photovoltaic and energy storage systems from hourly responses relying on human experience to minute-level closed loops driven by data and intelligent decision-making, significantly reducing operation and maintenance costs and improving economic benefits and power supply reliability.

[0007] In a preferred embodiment, this application can be further configured as follows: after real-time acquisition of multimodal data from the power generation side, spatiotemporal alignment and standardization fusion of the multimodal data to form a comprehensive energy dataset, wherein the multimodal data includes state parameters, energy storage side operating data, load side demand characteristics, and grid side economic signals, a multi-source data quality compensation and correction step is added, followed by the following steps: Real-time monitoring of sensor data anomalies; when an anomaly is detected, automatic switching to redundant sensors or virtual sensors based on data from nearby devices for recalculation, and marking the recalculation results with confidence labels; An adaptive filtering algorithm is used to smooth the energy storage working data to eliminate state jumps caused by sampling noise. At the same time, combined with ambient temperature feedforward compensation, the measurement deviation caused by temperature drift is corrected. Event detection and behavior pattern recognition are performed on the collected load demand characteristics to label high-power impact events and periodic fluctuation patterns. The labeling results are then embedded into the comprehensive energy dataset to provide explicit abrupt change feature inputs for subsequent prediction models.

[0008] In a preferred example, this application can be further configured as follows: the step of analyzing the comprehensive energy dataset based on a machine learning algorithm using a pre-set time-space dual-dimensional prediction model to output predicted power generation data for future periods includes the following steps: The encoder section employs a dynamic weighted graph attention network to automatically learn the spatial correlation weight distribution between power generation devices, identify spatial mismatch areas caused by shadow occlusion and uneven dust accumulation, and output spatial mismatch heat map feature vectors. The decoder part uses a bidirectional temporal modeling network to fuse the spatial feature vectors and historical sequences to generate power generation volatility predictions and load risk probability distributions. The volatility predictions are used to quantify the severity of power changes in future periods. A prediction confidence assessment branch is added to calculate the uncertainty range of the prediction value. When the confidence level is lower than the preset threshold, the prediction result is automatically downgraded to avoid high-risk prediction values ​​misleading the decision engine.

[0009] In a preferred embodiment, this application can be further configured as follows: in the step of inputting the predicted value into the intelligent decision engine and generating the optimal scheduling policy through policy optimization algorithm in a virtual simulation environment, the intelligent decision engine undergoes multi-objective hierarchical optimization and policy correction, the steps of which include: A top-level strategy network is constructed, with a long-term window as the decision range, to generate a power curve for energy storage scheduling. This power curve is used for global optimization. A mid-level strategy network is constructed, with an ultra-short-term time window as the rolling optimization step size. Based on the power generation forecast deviation and load change signal, the energy storage power curve for the current period is dynamically corrected to achieve rapid correction. A bottom-level strategy network is constructed to receive load impact events with high-frequency control cycles, generate load flexible scheduling instructions in real time, and adjust the energy storage charging and discharging rate synchronously.

[0010] In a preferred embodiment, this application can be further configured as follows: In the step of sending the strategy effectiveness verification report, actual execution data, and user feedback information to a cloud-based incremental learning platform, where the platform aggregates data from multiple systems based on a distributed collaborative learning mechanism and updates the parameters of the prediction model and decision engine online to achieve continuous evolution and scene adaptive optimization of the scheduling strategy, a user profile-driven scene adaptive optimization and model correction mechanism is added. The steps include: Based on historical electricity consumption data and feedback information, a user behavior profile is constructed. The profile includes electricity consumption habit tags and load elasticity coefficients, which are used to quantify the user's preference weights for economy, safety and environmental protection. The user profile is input into the cloud platform as prior knowledge, and online learning algorithms are used to incrementally update the prediction model and decision engine so that the model parameters continuously adapt to changes in user scenarios. When user feedback satisfaction falls below a preset threshold or a scene shift is detected, the model is automatically restarted and the policy is retrained to achieve scene-based adaptation of the scheduling policy.

[0011] In a preferred example, this application can be further configured as follows: In the step of constructing a user behavior profile based on historical electricity consumption data and feedback information, wherein the profile includes electricity consumption habit tags and load elasticity coefficients to quantify the user's preference weights for economy, safety, and environmental protection, the steps include the following: Multi-dimensional behavioral features are extracted from electricity consumption data, including load peak and valley period preferences, frequency of use of high-power equipment, sensitivity coefficient to electricity price fluctuations, and tolerance to unexpected power outages, forming a structured user feature vector. The system automatically identifies potential migration signs of user scenarios through time-series clustering algorithms. When a continuous deviation is detected in the load curve shape, electricity consumption period distribution, or equipment access type, a scenario migration warning is triggered, and a data buffering mechanism is initiated before the model is hot-rebooted. During the model hot restart process, a strategy of running the old and new strategies in parallel and a gradual switching strategy is adopted. The new strategy after retraining is run in shadow mode and the decision difference between it and the old strategy is compared. Only when the difference is stable and the user feedback indicators improve, the control weight is gradually transferred to the new strategy.

[0012] In a preferred embodiment, this application can be further configured to: in the step of converting the optimal scheduling strategy into a control instruction set conforming to the industry communication protocol via the northbound interface and issuing it to the power generation equipment, energy storage system, and load management terminal for execution, a virtual simulation verification and security fallback mechanism before strategy issuance is added, the steps including: The optimal scheduling strategy to be issued is injected into the virtual simulation platform to simulate the execution of the strategy and monitor key indicators to judge the battery. The virtual simulation platform includes a battery multiphysics coupling model, a power generation equipment degradation model and a load behavior model. If any abnormalities are detected during the simulation, such as excessive battery temperature field, excessive power generation abandonment rate, or increased number of load outages, the strategy will be immediately marked as unverified risk and rolled back to the previous safety baseline strategy. A two-way confirmation mechanism for command execution is added. After the policy is issued, the actual operating parameters of the device are read through the southbound interface, and the consistency between the issued value and the current value is compared. If the execution failure is found, the backup policy is automatically triggered and an early warning message is generated.

[0013] Secondly, the above-mentioned inventive objective of this application is achieved through the following technical solutions: An AI-based photovoltaic-storage energy dispatch optimization and recommendation device includes: a comprehensive energy dataset construction unit, used to collect multimodal data from the power generation side in real time, and to perform spatiotemporal alignment and standardization fusion of the multimodal data to form a comprehensive energy dataset, wherein the multimodal data includes state parameters, energy storage side operating data, load side demand characteristics, and grid side economic signals; The predictive power generation condition data calculation unit is used to pre-set a time-space dual-dimensional prediction model to analyze the comprehensive energy dataset based on a machine self-learning algorithm, so as to output the predicted power generation condition data for future periods, wherein the predicted power generation condition data includes the predicted power generation value, the predicted power generation volatility value, the predicted load value, and the load risk probability distribution. The optimal scheduling strategy generation unit is used to input the predicted value into the intelligent decision engine, perform strategy deduction in a virtual simulation environment through strategy optimization algorithm, and generate the optimal scheduling strategy, wherein the optimal scheduling strategy includes energy storage scheduling curve, power generation control value, and load scheduling instruction. The strategy effect verification report generation unit is used to convert the optimal scheduling strategy into a set of control instructions that conforms to the industry communication protocol through the northbound interface, and send it to the power generation equipment, energy storage system and load management terminal for execution. During the execution process, the unit collects the strategy execution deviation value in real time, calculates the residual between the deviation value and the predicted value, and generates a strategy effect verification report. The scheduling strategy optimization unit is used to send the strategy effect verification report, actual execution data and user feedback information to the cloud incremental learning platform. The platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, updates the parameters of the prediction model and decision engine online, and realizes the continuous evolution and scenario adaptive optimization of the scheduling strategy.

[0014] Thirdly, the above-mentioned objectives of this application are achieved through the following technical solutions: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aforementioned artificial intelligence-based photovoltaic-storage energy scheduling optimization recommendation method.

[0015] Fourthly, the above-mentioned objectives of this application are achieved through the following technical solutions: A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described artificial intelligence-based photovoltaic energy scheduling optimization recommendation method. Attached Figure Description

[0016] Figure 1This is a flowchart of an artificial intelligence-based photovoltaic energy dispatch optimization recommendation method in one embodiment of this application; Figure 2 This is a schematic diagram of a photovoltaic energy storage scheduling optimization and recommendation device based on artificial intelligence in one embodiment of this application; Figure 3 This is a schematic diagram of an electronic device according to an embodiment of this application.

[0017] Icon labels: 1. Integrated energy dataset construction unit; 2. Predictive power generation condition data calculation unit; 3. Optimal dispatch strategy generation unit; 4. Strategy effect verification report generation unit; 5. Dispatch strategy optimization unit. Detailed Implementation

[0018] The present application will be further described in detail below with reference to the accompanying drawings.

[0019] In one embodiment, such as Figure 1 As shown, this application discloses an artificial intelligence-based method for optimizing and recommending photovoltaic and energy storage scheduling, which specifically includes the following steps: S10: Real-time acquisition of multimodal data from the power generation side, spatiotemporal alignment and standardization fusion of the multimodal data to form a comprehensive energy dataset; Multimodal data from the generation side includes state parameters, energy storage side operating data, load side demand characteristics, and grid side economic signals. Specifically, in the rooftop photovoltaic-energy storage integrated system of a large commercial complex, edge computing nodes are deployed inside the photovoltaic inverters, energy storage converters, and floor distribution boxes. Nodes collect real-time data on the irradiance of the photovoltaic array, module surface temperature, and DC output power through sensor interfaces; obtain the state of charge, internal resistance, and module temperature field distribution of the energy storage batteries through the battery management system interface; collect load power curves and power factors of the air conditioning system, lighting system, and electric vehicle charging piles through smart meters; and read the daily time-of-use electricity price and demand price signals through the grid cloud platform interface. All data, after being aligned using a unified timestamp service, is standardized and integrated within the edge data lake to form a structured comprehensive energy dataset with an update frequency reaching the second level, providing a highly timely and multi-dimensional data foundation for subsequent forecasting.

[0020] S20: The pre-set time-space dual-dimensional prediction model analyzes the comprehensive energy dataset based on a machine self-learning algorithm to output predicted power generation data for future periods; The predicted power generation data includes predicted power generation, predicted power generation volatility, predicted load, and load risk probability distribution. Specifically, the comprehensive energy dataset is input into a pre-trained time-space dual-dimensional prediction model. The graph convolutional network in the model models the 30 strings of the rooftop photovoltaic array as graph nodes, automatically learning the spatial correlations between strings caused by shading and uneven dust accumulation, and identifying the mismatch pattern of the north-facing strings being shaded by nearby high-rise buildings in the afternoon. The long short-term memory network processes the power generation and load data of the past week along the time axis, capturing the differences in electricity consumption patterns between weekdays and weekends in the shopping mall. The model outputs predicted photovoltaic power generation and predicted power generation volatility for future periods, and predicts the probability distribution of peak risks formed by the overlap of a sudden load drop at the end of the evening business hours and concentrated charging at charging piles, providing accurate spatiotemporal prediction basis for scheduling decisions.

[0021] S30: Input the predicted value into the intelligent decision engine, and perform strategy deduction in a virtual simulation environment through the strategy optimization algorithm to generate the optimal scheduling strategy; The optimal scheduling strategy includes energy storage scheduling curves, power generation control values, and load scheduling instructions. Specifically, after receiving the predicted values, the intelligent decision engine initiates multi-objective optimization simulation. A reinforcement learning algorithm simulates 24-hour scheduling in a virtual simulation environment: during off-peak hours at night, the strategy instructs the energy storage system to charge at maximum power; during midday when photovoltaic power generation is high and load is moderate, the strategy prioritizes photovoltaic power consumption and moderately charges energy storage to reduce curtailment; before the evening load peak, the strategy initiates energy storage discharge in advance and flexibly commands the air conditioning system to operate at reduced load. The simulation generates an optimal scheduling strategy, including energy storage charging and discharging power curves, photovoltaic output control values, and flexible load scheduling instructions, maximizing electricity arbitrage profits while minimizing battery cycle losses and ensuring the risk of load outages approaches zero.

[0022] S40: The optimal scheduling strategy is converted into a set of control instructions that conforms to the industry communication protocol through the northbound interface and sent to the power generation equipment, energy storage system and load management terminal for execution. During the execution process, the strategy execution deviation value is collected in real time, and the residual value is calculated by comparing the deviation value with the predicted value to generate a strategy effect verification report. Specifically, the optimal scheduling strategy is converted into an industry-standard communication protocol instruction set via the northbound interface and sent to the photovoltaic inverter, energy storage converter, and load management terminal. Upon receiving the instruction, the inverter adjusts its MPPT operating point to match the photovoltaic control value, the energy storage converter executes charging and discharging according to the curve, and the load controller adjusts the air conditioning compressor frequency. During execution, the system collects the deviation between the actual power of each device and the instruction in real time, calculates the residual, and generates a strategy effectiveness verification report. For example, if the actual discharge power of the energy storage is lower than the instruction value due to a sudden increase in battery internal resistance, the verification report will mark the deviation during that period and trace it back to an abnormal battery health state, providing accurate feedback for subsequent optimization.

[0023] S50: The strategy effect verification report, actual execution data and user feedback information are sent to the cloud incremental learning platform. The platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, updates the parameters of the prediction model and decision engine online, and realizes the continuous evolution of the scheduling strategy and scenario adaptive optimization. Specifically, the strategy effectiveness verification report, actual execution data, and electricity user satisfaction evaluations from the shopping mall's property management were sent to the cloud-based incremental learning platform. Based on a federated learning mechanism, the platform aggregated data from the photovoltaic and energy storage systems of this shopping mall and 10 other similar commercial complexes nationwide, discovering that load forecasts during peak evening hours were generally overly optimistic. The platform updated the weight parameters of the Long Short-Term Memory network online to enhance its ability to capture random charging pile access behavior; simultaneously, it adjusted the economic weights in the reinforcement learning reward function and appropriately increased the priority of safety redundancy. The updated model was automatically distributed to edge nodes, making subsequent scheduling strategies more aligned with actual operating conditions and achieving continuous evolution in prediction accuracy and decision-making effectiveness. The update process employed differential privacy protection technology to ensure anonymization of data from all participants, complying with commercial data security standards. After receiving the new model, the edge nodes ran the old and new versions of the strategy in parallel through shadow mode, comparing key indicator differences and verifying stability before smoothly switching over. The system iterated the global model every 72 hours, forming a closed-loop optimization mechanism of "perception—decision—execution—feedback," effectively improving the efficiency of photovoltaic and energy storage collaboration and the consistency of user electricity experience.

[0024] In summary, compared with the traditional model of photovoltaic-storage scheduling that relies on fixed thresholds, human experience, and single-objective optimization, this application achieves a systematic leap in scheduling strategy by constructing a five-layer closed loop of "multi-source perception, multi-dimensional prediction, intelligent decision-making, reliable execution, and continuous evolution," which transforms the scheduling strategy from passive response to proactive prediction, from static solidification to dynamic adaptation, and from single-objective optimization to multi-objective collaboration. Specifically, multi-source data fusion and spatiotemporal alignment technologies significantly improve the accuracy of power generation and load forecasting, avoiding the risks of overcharging and discharging of energy storage or power outages caused by forecasting errors in traditional methods. The intelligent decision engine pre-simulates strategies in virtual simulations, ensuring that they are verified for both safety and economy before being deployed, preventing the direct execution of high-risk strategies. The cloud-based continuous learning mechanism enables the model to evolve with the accumulation of data from multiple systems, adapting to changes in different scenarios and avoiding the rigidity and failure of traditional models. The integrated hardware and software controller achieves a millisecond-level closed loop from perception to execution, supporting rapid response to second-level load impacts. This solution upgrades the scheduling of photovoltaic and energy storage systems from hourly responses relying on human experience to minute-level closed loops driven by data and intelligent decision-making, significantly reducing operation and maintenance costs, improving economic benefits and power supply reliability, and providing a replicable and evolvable intelligent scheduling paradigm for integrated photovoltaic and energy storage in various scenarios such as homes and industries.

[0025] In S10: Real-time acquisition of multimodal data from the power generation side, spatiotemporal alignment and standardization fusion of the multimodal data to form a comprehensive energy dataset, wherein the multimodal data includes state parameters, energy storage side operating data, load side demand characteristics, and grid side economic signals, followed by a multi-source data quality compensation and correction step, including the following steps: S11: Real-time monitoring of sensor data anomalies. When an anomaly is detected, it automatically switches to redundant sensors or virtual sensors based on data from nearby devices for recalculation, and marks the recalculation results with confidence labels. Specifically, in this embodiment, taking a large commercial complex rooftop photovoltaic system as an example, the edge node monitors in real time that the output value of the irradiance sensor of the photovoltaic string above the catering area is consistently more than 40% lower than that of the adjacent strings in the shopping area, and the value shows no daytime fluctuation, indicating that the sensor has failed due to oil fume deposition. The node automatically activates the virtual sensor supplementary calculation mechanism. Based on the irradiance data of two adjacent healthy strings, the tilt angle of the strings in the catering area, and the shading pattern over the past month, a virtual irradiance value is generated through a spatial interpolation algorithm, and the supplementary calculation result is labeled with the confidence level "virtual-medium". When loading data into the subsequent prediction model, the weight of this virtual data is automatically reduced by 30% in the calculation, avoiding deviations in power generation prediction during the midday catering peak due to single-point sensor failure, ensuring the integrity of the load dispatch strategy input, and significantly improving the system's robustness and continuous operation capability under sensor aging conditions.

[0026] S12: The adaptive filtering algorithm is used to smooth the energy storage working data to eliminate state jumps caused by sampling noise. At the same time, combined with ambient temperature feedforward compensation, the measurement deviation caused by temperature drift is corrected. Specifically, the current signal of the energy storage battery cluster in the underground equipment room of the commercial complex experiences high-frequency noise of ±10A due to frequent start-stop interference from nearby air conditioning chillers. This causes the estimated state of charge (SOC) value to jump by 4% within minutes, potentially triggering an alarm about insufficient backup power before the midday business peak. The edge node employs an adaptive filtering algorithm, combining battery cluster temperature field distribution data to dynamically adjust the filtering intensity to eliminate compressor electromagnetic interference, restoring a smooth and stable SOC curve. Simultaneously, the battery compartment temperature can reach 45℃ during summer midday, causing temperature drift in the current sensor. The node utilizes the temperature sensor within the battery compartment for feedforward compensation, correcting the measurement deviation by 0.6A. This process eliminates noise-induced SOC jumps, preventing false alarms from the protection system that could cause the energy storage system to prematurely exit backup mode, ensuring the accuracy of the assessment of load surge response capabilities during the midday business peak.

[0027] S13: Perform event detection and behavior pattern recognition on the collected load demand characteristics, label high-power impact events and periodic fluctuation patterns, and embed the labeling results into the comprehensive energy dataset to provide explicit mutation feature input for subsequent prediction models; Specifically, the edge node, through its load monitoring module, detected a sudden surge in the total load of air conditioning, lighting, and kitchen equipment in the dining area of ​​the commercial complex during the lunch period (11:30-13:00), from 800kW to 1800kW, with a rate of change exceeding 150kW / min. This was automatically marked as a "high-power surge event." Simultaneously, through two months of data learning, the system identified a periodic pattern of this surge occurring "concentrated during weekday lunch periods." Therefore, "lunchtime peak load surge" was embedded as a periodic fluctuation pattern label into the integrated energy dataset, and correlated with historical data on energy storage discharge depth and photovoltaic capacity for the same period. When the subsequent prediction model loads data, it automatically identifies this label and strengthens the load prediction weight for this period, reducing the lunchtime load prediction error from 12% to less than 5%. This provides a highly reliable decision-making basis for the energy storage system to switch to high-power discharge mode 30 minutes in advance and for flexible load pre-unloading, significantly improving the power supply reliability of the commercial complex during peak business hours.

[0028] In step S20: The pre-set time-space dual-dimensional prediction model analyzes the comprehensive energy dataset based on a machine self-learning algorithm to output predicted power generation data for future periods, the steps include the following: S201: The encoder section adopts a dynamic weighted graph attention network to automatically learn the spatial correlation weight distribution between power generation equipment, identify spatial mismatch areas caused by shadow occlusion and uneven dust accumulation, and output spatial mismatch heat map feature vectors. Specifically, in this embodiment, a photovoltaic array consisting of 30 strings is located on the roof of a large commercial complex, distributed across three areas: a dining area, a shopping area, and a parking shed. A dynamic weighted graph attention network treats each string as a graph node and automatically learns the spatial correlation between nodes through a self-attention mechanism: it identifies that the string in the dining area, due to long-term accumulation of grease and ash, has a lower correlation weight with the adjacent shopping area string, decreasing from 0.8 to 0.3; simultaneously, it detects that the string in the shopping area on the east side is shaded by nearby high-rise buildings in the afternoon, reducing its correlation weight with the unshaded string on the west side to 0.2. The network outputs a spatial mismatch heatmap, clearly marking areas with severe ash accumulation and shaded areas, providing explicit spatial mismatch feature vectors for subsequent prediction models. This allows the model to understand that the physical causes of "uneven power generation" are not merely numerical values, significantly improving the predictive sensitivity to localized shading.

[0029] S202: The decoder part uses a bidirectional temporal modeling network to fuse the spatial feature vector and historical sequence to generate power generation volatility prediction value and load risk probability distribution. The volatility prediction value is used to quantify the severity of power changes in future periods. Specifically, the decoder's bidirectional temporal modeling network integrates the spatial mismatch feature vector of S201 with the mall's historical power generation-load data from the past two weeks to identify the superposition pattern of a surge in air conditioning load in the catering area during lunchtime (11:30-13:00) and the impact of dust accumulation on photovoltaic output. The model generates a predicted value for power generation volatility: predicting that the power fluctuation amplitude of dust-covered strings will be significantly higher than that of clean strings when clouds pass through at midday; simultaneously, it outputs a load risk probability distribution, calculating that the probability of the load exceeding the threshold of 800kW during lunchtime is as high as 75%. The quantification of volatility and risk distribution provides a comparable basis for scheduling decisions, enabling the energy storage system to reserve a larger power margin in advance to cope with midday peaks, avoiding the shortcomings of traditional predictions that only output single-point values ​​and cannot assess the severity of fluctuations.

[0030] S203: Add a prediction confidence assessment branch to calculate the uncertainty range of the prediction value. When the confidence level is lower than the preset threshold, the prediction result downgrade mechanism is automatically triggered to avoid high-risk prediction values ​​misleading the decision engine. Specifically, the prediction model exhibits uncertainty in its forecasts of sudden weather events (such as severe afternoon convective weather). The confidence assessment branch calculates the uncertainty range of the predicted values ​​using the Monte Carlo Dropout method. When the confidence level of the predicted power generation of a dust-accumulated string falls below a preset threshold, a degradation mechanism is automatically triggered: the weight of the string's predicted value is reduced by 50% in the decision engine, and the decision relies more heavily on the prediction results of neighboring healthy strings and actual load-side measurements. This mechanism avoids misleading energy storage charging strategies with low-confidence, high-risk predictions, and prevents the risk of curtailment due to incorrect full charging of energy storage during severe convective weather caused by over-reliance on predictions, thus significantly improving the robustness and security of the overall dispatch.

[0031] This application, through the in-depth design of the prediction models S201-S203, upgrades the traditional "black box prediction" into an intelligent prediction system that is "explainable, quantifiable, and reliable." S201's spatial mismatch identification endows the model with physical understanding, accurately locating uneven power generation caused by dust accumulation and shading, providing a basis for differentiated scheduling. S202's volatility and risk probability quantification expands predictions from single-point values ​​to interval assessments, supporting the reservation of reasonable safety margins in scheduling strategies. S203's confidence degradation mechanism constructs a prediction reliability firewall, preventing low-quality predictions from propagating to the decision-making level and triggering cascading risks. The synergy of these three elements significantly enhances the scheduling system's adaptability to complex operating conditions (such as dust accumulation, shading, and sudden weather changes), reduces prediction errors, improves energy storage utilization, lowers the risk of load outages, and promotes the evolution of photovoltaic-storage scheduling from automation to intelligent autonomous operation.

[0032] In step S30: The predicted value is input into the intelligent decision engine, and the optimal scheduling policy is generated through policy optimization algorithms in a virtual simulation environment. This step involves multi-objective hierarchical optimization and policy correction of the intelligent decision engine, including the following steps: S301: Construct a top-level policy network, using a long-term window as the decision range, to generate a power curve for energy storage scheduling. The power curve is used for global optimization. Specifically, in this embodiment of the application, in a large commercial complex, the top-level strategy network generates a daily power curve for energy storage scheduling with a 24-hour decision-making scope. During the period from 00:00 to 06:00, when grid electricity prices are low, the strategy instructs the energy storage system to charge to 95% SOC at a maximum power of 300kW. During the period from 11:00 to 13:00, when solar power generation is high and the mall load is moderate, the strategy prioritizes solar power consumption and arranges for energy storage to charge at 150kW, reducing the curtailment rate to below 5%. During the peak business hours from 18:00 to 20:00, when the load surges, the strategy arranges for energy storage to discharge at 400kW to support air conditioning and lighting loads, reducing the need to purchase electricity from the high-priced grid. This power curve is globally optimized with the goal of maximizing the arbitrage profit of electricity prices throughout the day, avoiding the suboptimal overall economic performance caused by traditional methods that only focus on single-period benefits, and significantly improving the annualized rate of return of the solar-energy storage system.

[0033] S302: Construct a mid-level strategy network, using an ultra-short-term time window as the rolling optimization step size, and dynamically correct the energy storage power curve for the current period based on the power generation forecast deviation and load change signal to achieve rapid correction. At 12:30 PM, the mid-level strategy network detected that the load in the shopping mall's food and beverage area exceeded the predicted value by 200kW due to a sudden surge in customer traffic, while the photovoltaic output was 150kW lower than the predicted value due to cloud cover. The network immediately corrected the current energy storage power curve using a rolling optimization step of 15 minutes: the planned energy storage charging power was reduced from 150kW to 50kW, and 100kW of discharge power was released in advance to supplement the load gap; simultaneously, the flexible load management system was instructed to reduce the brightness of some non-critical lighting by 20% to quickly smooth out power fluctuations. This dynamic correction reduced the midday load peak by 10%, avoiding the risk of transformer overload, and achieving a minute-level closed loop of "prediction deviation - strategy adjustment - power balance," significantly reducing the risk of load outages or excessive energy storage discharge caused by prediction errors.

[0034] S303: Constructs an underlying strategy network to receive load impact events with a high-frequency control cycle, generate load flexible scheduling instructions in real time, and adjust the energy storage charging and discharging rate synchronously. Specifically, during the peak business hours at 18:15, the underlying strategy network received a load surge signal (power surge of 80kW) from the concentrated access of charging piles in the underground parking lot, with a control cycle of 1 second. The network immediately generated a flexible load dispatching instruction: within 5 seconds, it commanded the energy storage converter to increase the discharge power from 300kW to 380kW, simultaneously triggered the air conditioning unit frequency to decrease by 3Hz to reduce cooling power by 50kW, and instructed the charging pile management system to limit the power of all 20 charging guns to within 7kW. This second-level response achieved peak shaving and valley filling, preventing the surge from causing the 10kV incoming line switch to trip, and ensuring the normal power supply for the mall's operations. The underlying network, together with the top-level medium- and long-term strategy and the middle-level rolling optimization, formed a three-level "second-minute-time" collaboration, enabling the dispatching system to have both global economic optimization and local surge suppression capabilities.

[0035] This application, through a hierarchical strategy network design of S301-S303, upgrades the traditional single-time-scale scheduling strategy into a three-dimensional decision-making system with "second-level response, minute-level correction, and hour-level optimization." Top-level global optimization ensures optimal cross-time-period economics and avoids short-sighted decisions; mid-level rolling correction quickly suppresses the propagation of prediction errors, improving strategy robustness; and bottom-level high-frequency response effectively copes with load shocks, ensuring power supply reliability. The synergy of these three elements enables the photovoltaic-storage system in commercial complexes to maximize electricity arbitrage while avoiding equipment overload and power outage risks when facing complex operating conditions such as sudden changes in midday customer flow, evening peak hours, and concentrated charging pile access. This improves overall economics, reduces the risk of load outages, and promotes the evolution of photovoltaic-storage scheduling from automation to intelligent autonomous operation.

[0036] In step S40: Converting the optimal scheduling strategy into a control command set conforming to the industry communication protocol via the northbound interface and sending it to the power generation equipment, energy storage system, and load management terminal for execution, a virtual simulation verification and safety fallback mechanism is added before the strategy is sent. The steps include: S401: Inject the optimal scheduling strategy to be issued into the virtual simulation platform, simulate the execution of the strategy and monitor key indicators to judge the battery, wherein the virtual simulation platform includes a battery multiphysics coupling model, a power generation equipment degradation model and a load behavior model. Specifically, at 13:00 on summer business days, the intelligent decision engine generates the optimal scheduling strategy: energy storage is instructed to discharge at 500kW to support the load during the peak air conditioning cooling period from 14:00 to 16:00, while photovoltaic power is at full capacity and lighting is flexibly reduced by 20%. This strategy is then injected into a virtual simulation platform. The platform's built-in battery electrothermal coupling model simulates the temperature field distribution of the battery module under continuous 500kW discharge, the power generation equipment degradation model predicts the output decay curve of photovoltaic power at high temperatures, and the load behavior model reproduces the dynamic response of mall customer flow and air conditioning cooling demand. The simulation monitoring shows that the battery module temperature will approach the set threshold after 1.5 hours of discharge, posing a risk of overheating. This virtual simulation accurately identifies safety hazards before the strategy is issued, preventing high-risk strategies from directly affecting the physical system, essentially providing a "safety sandbox" for scheduling.

[0037] S402: During the simulation, if abnormalities such as excessive battery temperature field, excessive power generation abandonment rate, or increased number of load outages are detected, the strategy will be immediately marked as unverified risk and rolled back to the previous safety baseline strategy. Specifically, based on the S401 simulation results, the system detected an abnormal over-temperature in the battery temperature field and immediately marked the original strategy as "risk unverified," automatically reverting to the previous safe baseline strategy (energy storage discharge power reduced to 350kW, while simultaneously lowering the air conditioning unit frequency by 5Hz to compensate for the load reduction). This revert mechanism was completed before the actual strategy was issued at 14:00, taking only 2 minutes, and was imperceptible to the mall operator. After the strategy revert, virtual simulation retesting showed that the battery temperature stabilized within the safe range, and the load gap was filled by flexible load adjustment, without affecting the comfort of the business environment. This fallback mechanism ensures that every strategy is safely verified, eliminating the risk of "optimization failing and damaging equipment," and significantly improving the inherent safety level of system operation.

[0038] S403: Add a two-way confirmation mechanism for command execution. After the policy is issued, the actual operating parameters of the device are read through the southbound interface. The consistency between the issued value and the current value is compared. If the execution failure is found, the backup policy is automatically triggered and an early warning message is generated. Specifically, after the safety policy was issued at 14:00, the system read the actual operating parameters of the energy storage converter through the southbound interface and found that Energy Storage Unit 2 did not respond to the discharge command due to an intermittent communication module failure, with an actual discharge power of only 320kW (lower than the issued value of 350kW). The two-way confirmation mechanism immediately triggered an early warning and automatically activated the backup strategy: instructing Energy Storage Units 1 and 3 to each increase the discharge power by 15kW to compensate for the shortfall, and simultaneously notifying maintenance personnel to inspect Unit 2. This mechanism completed the activation of the backup strategy and alarm push within 30 seconds, and the power supply to the shopping mall was not affected. Two-way confirmation ensures that "instruction is not equal to execution," achieving full controllability of the strategy implementation process, forming a complete closed loop of "issuance-confirmation-correction-backup," and ensuring the reliable execution of dispatch commands and the continuous and stable operation of the system.

[0039] In step S50: The strategy effectiveness verification report, actual execution data, and user feedback information are sent to the cloud-based incremental learning platform. This platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, and updates the parameters of the prediction model and decision engine online to achieve continuous evolution and scenario-adaptive optimization of the scheduling strategy. In this step, a user profile-driven scenario-adaptive optimization and model correction mechanism is added. The steps include: S501: Based on historical electricity consumption data and feedback information, construct a user behavior profile. The profile includes electricity consumption habit tags and load elasticity coefficients, which are used to quantify the user's preference weights for economy, safety and environmental protection. Specifically, based on the commercial complex's historical electricity consumption data from the past year and property management feedback, the system automatically constructs user behavior profiles: identifying that during the summer weekday period from 11:00 to 13:00, the air conditioning load in the catering area accounts for as much as 45%, and the property management is highly sensitive to time-of-use electricity prices, marking this as "midday electricity price sensitive type"; during the winter period from 18:00 to 20:00, lighting and heating loads are concentrated, and the property management prioritizes power supply reliability over economy, marking this as "evening safety priority type"; on weekends and holidays, the load curve is flat throughout the day, and the load elasticity coefficient is high, marking this as "elastic adjustable type". The profile quantifies the user's preference weights for economy, safety, and environmental protection (set to 0.7 for economy and 0.3 for safety during summer midday in this embodiment), providing interpretable prior knowledge for subsequent incremental updates to the model, avoiding the problem of misalignment between strategies and actual user needs caused by the "one-size-fits-all" approach of traditional models.

[0040] S502: Input the user profile as prior knowledge into the cloud platform, and use online learning algorithms to incrementally update the prediction model and decision engine so that the model parameters continuously adapt to changes in user scenarios. Specifically, the cloud platform embeds the user profiles built by S501 as prior knowledge into the online learning algorithm: during incremental updates, the loss function of the time-space prediction model is weighted, amplifying the penalty term for midday load prediction error by 1.5 times, forcing the model to more accurately learn the correlation between air conditioning load and temperature and pedestrian flow; the reward function of the decision engine is reinforced, dynamically increasing the weight of midday economic efficiency, making the energy storage charging strategy more inclined to deep charging during low electricity prices (00:00-06:00) and maximizing arbitrage during high-price periods at midday. After three months of online learning, the midday load prediction error decreased from 8% to 3%, energy storage arbitrage revenue increased by 12%, and the model parameters continuously adapted to the seasonal changes in shopping mall electricity consumption, achieving a smooth evolution from a "general model" to a "scenario-specific model," significantly improving strategy fit and user satisfaction.

[0041] S503: When user feedback satisfaction is lower than the preset threshold or a scene shift is detected, the model is automatically restarted and the policy is retrained to achieve scene-based adaptation of the scheduling policy. Specifically, in this embodiment, due to business format adjustments, a commercial complex converted some areas into a 24-hour e-sports arena, causing the nighttime load percentage to surge from 5% to 25%, resulting in a continuous shift in load distribution during different time periods. The system detected that the average nighttime load exceeded twice the historical baseline for seven consecutive days, and the load elasticity coefficient decreased significantly, automatically triggering a "scene migration warning." The cloud platform initiated a model hot restart process: the retrained new strategy was run in parallel in shadow mode, comparing the differences between the old and new strategies in key instructions such as nighttime energy storage charging and discharging, and load adjustment. It was found that the new strategy more accurately balanced nighttime e-sports load and energy storage backup power demand without increasing battery cycle losses. The control weights were then gradually switched from the old strategy to the new strategy, a process lasting two hours, during which power supply reliability indicators were monitored in real time to ensure zero power outages. This mechanism achieves a seamless and smooth transition during scene migration, avoiding the risk of power outages caused by traditional "downtime upgrades," and ensuring the stability of the energy system during the commercial complex's business format upgrade process.

[0042] This application upgrades the traditional fixed cloud model to a refined and continuously evolving system that is "personalized" by adding user profile-driven scenario adaptation mechanisms (S501-S503) after S50. S501 automatically constructs interpretable user behavior profiles, accurately quantifies user preference weights in different scenarios, and solves the problem of misalignment between strategies and actual user needs. S502 uses profiles as prior knowledge to guide incremental model updates, enabling prediction and decision parameters to continuously adapt to seasonal and business-related changes, improving strategy fit and economy. S503's scenario migration detection and hot restart mechanism achieves a zero-risk smooth transition in the model upgrade process, ensuring continuous energy supply. The synergy of these three mechanisms enables the photovoltaic-storage scheduling system to possess autonomous intelligence of "understanding users - adapting to scenarios - seamless evolution," increasing user satisfaction by more than 20%, shortening the strategy iteration cycle from monthly to weekly, and promoting the transformation of photovoltaic-storage energy management from an "automated execution" to an "intelligent service" paradigm.

[0043] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0044] In one embodiment, an AI-based photovoltaic-storage energy scheduling optimization recommendation device is provided, which corresponds one-to-one with the AI-based photovoltaic-storage energy scheduling optimization recommendation method described in the above embodiments. For example... Figure 2 As shown, the AI-based photovoltaic-storage energy dispatch optimization recommendation device includes: a comprehensive energy dataset construction unit 1, which is used to collect multimodal data from the power generation side in real time, and to perform spatiotemporal alignment and standardization fusion of the multimodal data to form a comprehensive energy dataset. The multimodal data includes state parameters, energy storage side working data, load side demand characteristics, and grid side economic signals. The predictive power generation condition data calculation unit 2 is used to pre-set a time-space dual-dimensional prediction model to analyze the comprehensive energy dataset based on a machine self-learning algorithm, so as to output the predicted power generation condition data for future periods, wherein the predicted power generation condition data includes the predicted power generation value, the predicted power generation volatility value, the predicted load value, and the load risk probability distribution. The optimal scheduling strategy generation unit 3 is used to input the predicted value into the intelligent decision engine, perform strategy deduction in a virtual simulation environment through strategy optimization algorithm, and generate the optimal scheduling strategy, wherein the optimal scheduling strategy includes energy storage scheduling curve, power generation control value, and load scheduling instruction. The strategy effect verification report generation unit 4 is used to convert the optimal scheduling strategy into a set of control instructions that conforms to the industry communication protocol through the northbound interface, and send it to the power generation equipment, energy storage system and load management terminal for execution. During the execution process, the strategy execution deviation value is collected in real time, and the residual value is calculated by comparing the deviation value with the predicted value to generate a strategy effect verification report. The scheduling strategy optimization unit 5 is used to send the strategy effect verification report, actual execution data and user feedback information to the cloud incremental learning platform. The platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, updates the parameters of the prediction model and decision engine online, and realizes the continuous evolution and scenario adaptive optimization of the scheduling strategy.

[0045] Specific limitations regarding the AI-based photovoltaic-storage energy dispatch optimization recommendation device can be found in the limitations of the AI-based photovoltaic-storage energy dispatch optimization recommendation method described above, and will not be repeated here. Each module in the aforementioned AI-based photovoltaic-storage energy dispatch optimization recommendation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the electronic device, or stored in the memory of the electronic device as software, so that the processor can call and execute the corresponding operations of each module.

[0046] In one embodiment, an electronic device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, this electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the database. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an artificial intelligence-based photovoltaic-storage energy scheduling optimization recommendation method.

[0047] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Multimodal data from the power generation side is collected in real time, and the multimodal data is spatiotemporally aligned and standardized and fused to form a comprehensive energy dataset. The multimodal data includes state parameters, energy storage side operating data, load side demand characteristics, and grid side economic signals. The pre-set time-space dual-dimensional prediction model analyzes the comprehensive energy dataset based on a machine self-learning algorithm to output predicted power generation operating conditions data for future periods, wherein the predicted power generation operating conditions data includes predicted power generation, predicted power generation volatility, predicted load, and load risk probability distribution. The predicted values ​​are input into the intelligent decision engine, and the strategy optimization algorithm is used to perform strategy deduction in a virtual simulation environment to generate the optimal scheduling policy, wherein the optimal scheduling policy includes energy storage scheduling curve, power generation control value, and load scheduling instruction. The optimal scheduling strategy is converted into a set of control instructions that conforms to the industry communication protocol through the northbound interface and sent to the power generation equipment, energy storage system and load management terminal for execution. During the execution process, the strategy execution deviation value is collected in real time, and the residual value is calculated by comparing the deviation value with the predicted value to generate a strategy effect verification report. The strategy effectiveness verification report, actual execution data, and user feedback information are sent to the cloud-based incremental learning platform. The platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, updates the parameters of the prediction model and decision engine online, and realizes continuous evolution and scenario adaptive optimization of the scheduling strategy.

[0048] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Multimodal data from the power generation side is collected in real time, and the multimodal data is spatiotemporally aligned and standardized and fused to form a comprehensive energy dataset. The multimodal data includes state parameters, energy storage side operating data, load side demand characteristics, and grid side economic signals. The pre-set time-space dual-dimensional prediction model analyzes the comprehensive energy dataset based on a machine self-learning algorithm to output predicted power generation operating conditions data for future periods, wherein the predicted power generation operating conditions data includes predicted power generation, predicted power generation volatility, predicted load, and load risk probability distribution. The predicted values ​​are input into the intelligent decision engine, and the strategy optimization algorithm is used to perform strategy deduction in a virtual simulation environment to generate the optimal scheduling policy, wherein the optimal scheduling policy includes energy storage scheduling curve, power generation control value, and load scheduling instruction. The optimal scheduling strategy is converted into a set of control instructions that conforms to the industry communication protocol through the northbound interface and sent to the power generation equipment, energy storage system and load management terminal for execution. During the execution process, the strategy execution deviation value is collected in real time, and the residual value is calculated by comparing the deviation value with the predicted value to generate a strategy effect verification report. The strategy effectiveness verification report, actual execution data, and user feedback information are sent to the cloud-based incremental learning platform. The platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, updates the parameters of the prediction model and decision engine online, and realizes continuous evolution and scenario adaptive optimization of the scheduling strategy.

[0049] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0050] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0051] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An artificial intelligence-based light storage energy scheduling optimization recommendation method, characterized in that, The method comprises the steps of: collecting multi-modal data of a power generation side in real time, performing space-time alignment and standardized fusion on the multi-modal data to form a comprehensive energy data set, wherein the multi-modal data comprises state parameters, energy storage side working data, load side demand characteristics and power grid side economic signals; A pre-set time-space two-dimensional prediction model analyzes the comprehensive energy data set based on a machine self-learning algorithm to output predicted power generation working condition data of a future period, wherein the predicted power generation working condition data comprises a power generation amount prediction value, a power generation fluctuation rate prediction value, a load prediction value and a load risk probability distribution; The prediction value is input into an intelligent decision engine to perform strategy deduction in a virtual simulation environment through a strategy optimization algorithm to generate an optimal scheduling strategy, wherein the optimal scheduling strategy comprises an energy storage scheduling curve, a power generation control value and a load scheduling instruction; The optimal scheduling strategy is converted into a control instruction set conforming to an industry communication protocol through a northbound interface and is issued to power generation equipment, an energy storage system and a load management terminal for execution, and a strategy execution deviation value is collected in real time during the execution process, a residual error calculation is performed on the deviation value and the prediction value, and a strategy effect verification report is generated; The strategy effect verification report, actual execution data and user feedback information are sent to a cloud incremental learning platform, the platform aggregates data of multiple systems based on a distributed collaborative learning mechanism, and parameters of the prediction model and the decision engine are updated online to realize continuous evolution and scene adaptive optimization of the scheduling strategy.

2. The method of claim 1, wherein the method further comprises: After the step of collecting multi-modal data of a power generation side in real time, performing space-time alignment and standardized fusion on the multi-modal data to form a comprehensive energy data set, wherein the multi-modal data comprises state parameters, energy storage side working data, load side demand characteristics and power grid side economic signals, a step of multi-source data quality compensation and correction is added, and the steps comprise: Abnormal situations of sensor data are monitored in real time, when data anomalies are detected, redundant sensors or virtual sensors based on adjacent equipment data are automatically switched to perform compensation calculation, and the compensation calculation result is marked with a confidence label; An adaptive filtering algorithm is used to smooth the energy storage working data to eliminate state jumps caused by sampling noise, and environmental temperature feedforward compensation is combined to correct measurement deviations caused by temperature drift; Event detection and behavior pattern recognition are performed on the collected load demand characteristics, high-power impact events and periodic fluctuation patterns are marked, and the marking results are embedded in the comprehensive energy data set to provide explicit mutation feature input for the subsequent prediction model.

3. The method of claim 2, wherein, In the step of analyzing the comprehensive energy data set based on a machine self-learning algorithm by a pre-set time-space two-dimensional prediction model to output predicted power generation working condition data of a future period, the steps comprise: A dynamic weight map attention network is used in the encoder part to automatically learn the spatial correlation weight distribution between power generation equipment, identify spatial mismatch areas caused by shadow shielding and uneven dust accumulation, and output a spatial mismatch thermal map feature vector; The decoder part adopts a bidirectional time sequence modeling network to fuse the spatial feature vector and historical sequence, generate power fluctuation rate prediction value and load risk probability distribution, and the fluctuation rate prediction value is used to quantify the intensity of power change in future period; A prediction confidence evaluation branch is added to calculate the uncertainty range of the prediction value. When the confidence is lower than the preset threshold, the prediction result degradation mechanism is automatically triggered to avoid misleading the decision engine with high-risk prediction values.

4. The method of claim 3, wherein, In the step of inputting the prediction value into the intelligent decision engine and generating the optimal scheduling strategy through strategy optimization algorithm in the virtual simulation environment, the intelligent decision engine is subjected to multi-objective hierarchical optimization and strategy correction, including the following steps: A top strategy network is constructed to generate a power curve for energy storage scheduling with a long time window as the decision range, and the power curve is used for global optimization; A middle strategy network is constructed to dynamically correct the energy storage power curve in the current period according to the generation prediction deviation and load mutation signal, and realize rapid correction; A bottom strategy network is constructed to receive load impact events at a high-frequency control period, generate load flexible scheduling instructions in real time, and adjust the charging and discharging rate of energy storage synchronously.

5. The method of claim 1, wherein, In the step of sending the strategy effect verification report, actual execution data and user feedback information to the cloud incremental learning platform, the platform aggregates data from multiple systems based on a distributed collaborative learning mechanism, and updates the parameters of the prediction model and the decision engine online to realize continuous evolution and scene adaptive optimization of the scheduling strategy, a user portrait driven scene adaptive optimization and model correction mechanism is added, including the following steps: Based on historical electricity consumption data and feedback information, a user behavior feature portrait is constructed, which includes electricity habit tags and load elasticity coefficients, which are used to quantify the user's preference weights for economy, safety and environmental protection; The user portrait is input as prior knowledge into the cloud platform, and the prediction model and the decision engine are updated incrementally using online learning algorithm, so that the model parameters continuously adapt to the changes of user scenarios; When the user feedback satisfaction is lower than the preset threshold or the scene migration is detected, the model hot restart and strategy retraining are automatically triggered to realize the scene adaptive optimization of the scheduling strategy.

6. The method of claim 5, wherein the method further comprises: In the step of constructing a user behavior feature portrait based on historical electricity consumption data and feedback information, the portrait includes electricity habit tags and load elasticity coefficients, which are used to quantify the user's preference weights for economy, safety and environmental protection, including the following steps: Multi-dimensional behavior features are extracted from electricity consumption data, including load peak and valley period preference, high-power equipment usage frequency, sensitivity coefficient to electricity price fluctuations, and tolerance to unexpected power failure, forming a structured user feature vector; Through time sequence clustering algorithm, the potential migration signs of user scenarios are automatically identified. When the load curve shape, electricity consumption period distribution or equipment access type deviates continuously, the scene migration warning is triggered, and the data buffering mechanism before model hot restart is started; During the model hot restart process, the new and old strategies are run in parallel and the gradual switching strategy is used. The retrained new strategy is run in shadow mode. The decision differences between the new strategy and the old strategy are compared. Only when the differences are stable and the user feedback indicators improve, the control weight is gradually transferred to the new strategy.

7. The method of claim 1, wherein, In the step of converting the optimal scheduling strategy into a control instruction set conforming to the industry communication protocol through the northbound interface, and issuing it to the power generation equipment, energy storage system and load management terminal for execution, a virtual deduction verification and safety bottom mechanism before strategy issuance is added, including the following steps: Injecting the optimal scheduling strategy to be issued into a virtual simulation platform to simulate the execution of the strategy and monitor key indicators to determine the battery, wherein the virtual simulation platform includes a battery multi-physical field coupling model, a power generation equipment degradation model and a load behavior model; During the deduction process, if the battery temperature field is overheated, the power generation abandonment rate exceeds the standard or the number of load power outages increases, etc., the strategy is immediately marked as risk unverified and rolled back to the last safe baseline strategy; A bidirectional confirmation mechanism for instruction execution is added. After the strategy is issued, the actual running parameters of the equipment are read through the southbound interface to compare the consistency of the issued value and the current value. If the execution fails, the standby strategy is automatically triggered and a warning information is generated.

8. An artificial intelligence-based photovoltaic and energy storage energy scheduling optimization recommendation device applied to the artificial intelligence-based photovoltaic and energy storage energy scheduling optimization recommendation method of any one of claims 1 to 7, characterized in that, The device comprises: An integrated energy data set construction unit (1) for real-time acquisition of multi-modal data on the power generation side, time-space alignment and standardized fusion of the multi-modal data to form an integrated energy data set, wherein the multi-modal data includes state parameters, energy storage side working data, load side demand characteristics and grid side economic signals; A predicted power generation working condition data calculation unit (2) for pre-setting a time-space two-dimensional prediction model to analyze the integrated energy data set based on a machine self-learning algorithm to output predicted power generation working condition data for a future period, wherein the predicted power generation working condition data includes predicted power generation value, predicted power generation fluctuation rate, predicted load value and load risk probability distribution; An optimal scheduling strategy generation unit (3) for inputting the predicted value into an intelligent decision engine to generate an optimal scheduling strategy through strategy optimization algorithm in a virtual simulation environment, wherein the optimal scheduling strategy includes energy storage scheduling curve, power generation control value and load scheduling instruction; A strategy effect verification report generation unit (4) for converting the optimal scheduling strategy into a control instruction set conforming to the industry communication protocol through the northbound interface, issuing it to the power generation equipment, energy storage system and load management terminal for execution, and real-time acquisition of strategy execution deviation value during the execution process, residual calculation of the deviation value and the predicted value to generate a strategy effect verification report; A scheduling strategy optimization unit (5) for sending the strategy effect verification report, actual execution data and user feedback information to a cloud incremental learning platform, which aggregates data of multiple systems based on a distributed collaborative learning mechanism to update parameters of the prediction model and the decision engine online, realizing continuous evolution and scene adaptive optimization of the scheduling strategy.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor implements the steps of the artificial intelligence-based photovoltaic and energy storage energy scheduling optimization recommendation method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the steps of the artificial intelligence-based photovoltaic and energy storage energy scheduling optimization recommendation method according to any one of claims 1 to 7.

Images