Artificial intelligence-driven dynamic optimization scheduling system for photovoltaic absorption type battery charging and swapping micro-grid

Abstract

The invention relates to the technical field of micro-grids, in particular to an artificial intelligence-driven photovoltaic absorption type charging micro-grid dynamic optimization scheduling system. Comprising a data acquisition and preprocessing unit; the AI-driven dynamic optimization scheduling unit takes the confidence interval as an input state variable of the improved depth deterministic strategy gradient algorithm; a photovoltaic absorption-charging and battery swap cooperative control unit; and a state monitoring and feedback unit. According to the invention, through a dynamic optimization scheduling unit driven by AI, a risk perception reinforcement learning module (which adds confidence interval width to improve an algorithm state variable of DDPG) and a strategy adaptive adjustment module (which calculates a dynamic risk weight into a reward function according to the confidence interval width, and preferentially extracts high-uncertainty experience in an experience pool) are used; the problem of lack of prediction risk adaptation capability in the prior art is solved, the scheduling strategy dynamically adjusts the reward function and the exploration rate along with the photovoltaic prediction risk, the response flexibility is improved, and the problem of supply and demand imbalance is reduced.

Description

Technical Field

[0001] This invention relates to the field of microgrid technology, and more specifically, to an artificial intelligence-driven dynamic optimization scheduling system for photovoltaic-integrated charging and swapping microgrids. Background Technology

[0002] With the large-scale application of photovoltaic energy in the charging and swapping field, photovoltaic-integrated charging and swapping microgrids have become a key carrier for balancing clean energy utilization and charging service demand. In the operation of such microgrids, photovoltaic output is easily affected by meteorological factors such as sunlight and temperature, resulting in fluctuations, while charging and swapping loads exhibit randomness due to differences in user habits. This dynamic imbalance often leads to problems such as unstable microgrid bus voltage, increased photovoltaic curtailment rates, or inability to meet charging demands in a timely manner. Therefore, developing an optimized scheduling scheme that can adapt to the dynamic characteristics of the microgrid and achieve precise matching between photovoltaic output and charging and swapping loads has become a core requirement in this field.

[0003] In the existing technology, relevant patents have explored the field of microgrid collaborative control and regulation capability assessment. For example, Chinese patent CN202510619479.5 discloses a collaborative control method for multiple microgrids participating in grid interaction at the park level. This method includes: acquiring real-time data from the grid platform and multiple microgrids, where each microgrid contains distributed power sources, energy storage, and loads; determining the grid's grid-connected or off-grid operation status based on the data; performing energy management analysis based on the operation status and real-time data and generating control commands; implementing frequency and voltage stability control of the underlying equipment of the multiple microgrids according to the control commands, and simultaneously formulating energy trading strategies. This invention achieves stable microgrid control and improves the economy, stability, and operating efficiency of multi-microgrid systems. Chinese patent CN202510586371.0 discloses a microgrid regulation capability evaluation system and method based on photovoltaic-storage-charging collaborative interaction, including five modules: multi-source data fusion, dynamic evaluation, collaborative optimization, digital twin verification, and output control. Each module establishes a signal connection in sequence. It adopts hybrid spatiotemporal data fusion technology to reduce synchronization error, reduces computational complexity through dynamic evaluation-optimization joint solution, and reduces computational load by using an event-driven rolling optimization mechanism.

[0004] Despite the design advantages of the above technical solutions, they also have the following technical defects: First, they lack a multi-timescale prediction and uncertainty quantification mechanism for photovoltaic output. Chinese patent CN202510619479.5 relies solely on real-time grid data for energy management and does not design adaptive prediction logic for the differences in photovoltaic output at different time windows (such as short-term fluctuations at the hourly level and medium-to-long-term changes at the daily level), thus failing to meet the accuracy requirements of short-term charging load matching and medium-to-long-term energy storage planning. Although the dynamic evaluation system of Chinese patent CN202510586371.0 is linked to photovoltaic, energy storage, and charging data, it does not quantify the deviation range or reliability of photovoltaic output prediction results, making it difficult to predict in advance the scheduling risks that may be caused by inaccurate predictions. Secondly, it lacks the adaptive adjustment capability of the scheduling strategy based on predicted risks: After generating control commands, Chinese patent CN202510619479.5 does not adjust parameters according to the actual fluctuations in photovoltaic output during subsequent operation, and cannot cope with power shortages or curtailment caused by prediction deviations; the event-driven optimization of Chinese patent CN202510586371.0 only uses output fluctuation rate and load mutation rate as trigger conditions, and does not adjust the optimization priority according to the risk level of the prediction results (such as the prediction credibility during high fluctuation periods). The scheduling strategy lacks flexibility in responding to the dynamic risks of the microgrid and is difficult to adapt to the complex needs of photovoltaic-integrated charging and swapping scenarios. In view of this, we propose an artificial intelligence-driven dynamic optimization scheduling system for photovoltaic-integrated charging and swapping microgrids. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-driven dynamic optimization scheduling system for photovoltaic-integrated charging and swapping microgrids to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention aims to provide an artificial intelligence-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids, comprising: The data acquisition and preprocessing unit collects real-time operating data related to photovoltaic array output, charging and swapping load, microgrid bus electrical parameters and electricity price through distributed sensors, smart meters and microgrid monitoring system. The unit performs noise reduction and standardization preprocessing on the collected data, generates a standardized dataset and stores it in the time-series real-time database. The AI-driven dynamic optimization scheduling unit is based on a standardized dataset. It constructs a photovoltaic power output prediction model through a multi-timescale long short-term memory network, outputting the photovoltaic power output prediction value and its confidence interval. The confidence interval is used as the input state variable of the improved deep deterministic strategy gradient algorithm. A dynamic risk weight term based on the confidence interval width is introduced to adaptively adjust the reward function and exploration rate parameters, generating a dynamic scheduling strategy that adapts to the dynamic characteristics of the microgrid. Photovoltaic accommodation - charging and swapping协同control unit, the photovoltaic accommodation - charging and swapping协同control unit receives dynamic scheduling strategies, performs orderly charging control on charging piles, and performs energy storage scheduling on the energy storage system supporting the swapping station, so as to achieve dynamic matching of photovoltaic output and charging and swapping loads; Status monitoring and feedback unit, the status monitoring and feedback unit collects real - time operating status data of the microgrid in real time, compares the deviation with the predicted value in the dynamic scheduling strategy, and when the deviation exceeds the preset deviation threshold, triggers the AI - driven dynamic optimization scheduling unit to recalculate and update the scheduling strategy iteratively, forming a closed - loop scheduling mechanism.

[0007] As a further improvement of this technical solution, the data acquisition and pre - processing unit includes a data acquisition module and a data pre - processing and storage module, where: The data acquisition module collects photovoltaic panel temperature, light intensity, and photovoltaic array output power through distributed sensors, collects real - time power, current, and electricity price data of charging and swapping loads through smart meters, and collects microgrid bus voltage, current, total active power, and reactive power through the microgrid monitoring system; The data pre - processing and storage module uses the 3σ criterion to剔除异常值process various raw data obtained by the data acquisition module, maps the denoised data to a unified numerical interval through the min - max normalization method to generate a standardized data set; and stores this standardized data set in the time - series real - time database according to the dimensions of "data type - acquisition time - device number" to support the subsequent AI - driven dynamic optimization scheduling unit to quickly query and call data.

[0008] As a further improvement of this technical solution, the AI - driven dynamic optimization scheduling unit includes a multi - time - scale prediction module, an uncertainty quantification module, a risk - aware reinforcement learning module, and a policy adaptive adjustment module, where: The multi - time - scale prediction module realizes short - term and medium - long - term prediction of photovoltaic output based on a multi - time - window long - short - term memory network and outputs a prediction confidence interval; The uncertainty quantification module calculates the prediction confidence interval of photovoltaic output according to the prediction feature sequence; The risk - aware reinforcement learning module introduces the confidence interval information as an input state variable into the improved deep deterministic policy gradient algorithm; The policy adaptive adjustment module dynamically adjusts the reward function weight and exploration rate parameter of the reinforcement learning algorithm according to the confidence interval to realize the generation of a dynamic scheduling strategy.

[0009] As a further improvement of this technical solution, the process of the multi - time - scale prediction module outputting a prediction confidence interval includes the following steps: S21.1 Set short-term and medium-to-long-term time windows, and construct short-term prediction sub-networks and medium-to-long-term prediction sub-networks respectively. Both short-term and medium-to-long-term prediction sub-networks adopt the long short-term memory network structure. S21.2 Input short-term photovoltaic power output historical data from the standardized dataset into the short-term prediction subnetwork, and input medium- and long-term photovoltaic power output historical data and corresponding meteorological data from the standardized dataset into the medium- and long-term prediction subnetwork. Minimize the error between the predicted value and the actual value through the gradient descent method, and iterate until the error of the validation set converges and stabilizes. S21.3. The prediction outputs of the short-term prediction sub-network and the medium- and long-term prediction sub-network are weighted and fused through the attention weight fusion mechanism. During the fusion, the weight ratio of the outputs of the short-term prediction sub-network and the medium- and long-term prediction sub-network is dynamically allocated according to the microgrid's requirements for prediction accuracy at different time periods to obtain the preliminary photovoltaic power output prediction value. S21.4 Transmit the preliminary photovoltaic power output forecast after fusion to the uncertainty quantification module.

[0010] As a further improvement to this technical solution, the process of the uncertainty quantification module calculating the confidence interval for photovoltaic power output prediction includes the following steps: S22.1. Embed the Monte Carlo dropout method in the long short-term memory network of the multi-timescale prediction module to ensure that the dropout layer in the network training phase continues to be effective in the prediction phase. S22.2. The input data of the multi-timescale prediction module is passed through a long short-term memory network that embeds the Monte Carlo dropout method and keeps the dropout layer active for multiple independent forward propagations to obtain multiple parallel prediction output values. S22.3 Statistical analysis is performed on the output values ​​of multiple parallel predictions to calculate the final photovoltaic power output prediction value. and variance ; S22.4, combined with confidence coefficient (Selected based on the operational stability requirements of microgrid equipment), construct the confidence interval for photovoltaic power output prediction, and calculate the width of the confidence interval. Confidence interval and width The output is sent to the risk perception reinforcement learning module.

[0011] As a further improvement to this technical solution, the process of embedding confidence interval information into the risk perception reinforcement learning module includes the following steps: S23.1. Based on the original state variables of the improved deep deterministic policy gradient algorithm, a new confidence interval width is added. As state variables, they expand the state input dimension of the algorithm; S23.2, Based on the width of the confidence interval Width of the historical maximum confidence interval The ratio is used to classify uncertainty levels and set the exploration rate. The range of values, where a first threshold is set. Second threshold ,and : when hour, Take the lower range of values; when hour, Take the middle range of values; when When, take the higher value range; S23.3. Input the expanded complete state variables into the actor network of the algorithm, so that the reinforcement learning agent can directly perceive the uncertainty of prediction during the decision-making process.

[0012] As a further improvement to this technical solution, the process of the strategy adaptive adjustment module optimizing algorithm parameters and generating scheduling strategies includes the following steps: S24.1 Confidence interval width based on the output of the uncertainty quantization module Width of the historical maximum confidence interval Calculate dynamic risk weights ; S24.2, Dynamic risk weights Introducing a reward function to form a multi-objective technical optimization reward function ; S24.3. Introduce a priority sampling rule and set a threshold for the empirical replay pool of the improved deep deterministic strategy gradient algorithm. ,right Scheduling experience is assigned a higher sampling weight, and the actor-evaluator network parameters of this type of experience-based update algorithm are extracted first; among which... satisfy ; S24.4 Based on the output of the optimized and improved deep deterministic strategy gradient algorithm, combined with the rated power of the microgrid photovoltaic inverter, the charge and discharge rate limit of the energy storage system, and the capacity constraint of the charging pile, a dynamic scheduling strategy including photovoltaic power distribution instructions, energy storage charge and discharge power instructions, and charging pile power scheduling instructions is generated and output to the photovoltaic consumption-charging and swapping collaborative control unit.

[0013] As a further improvement to this technical solution, the photovoltaic consumption-charging / swapping coordinated control unit includes a strategy receiving and parsing module, a charging terminal demand management module, a charging pile orderly control module, an energy storage system scheduling module, and a supply and demand matching monitoring module, wherein: The policy receiving and parsing module is used to receive the dynamic scheduling policy output by the policy adaptive adjustment module, and disassemble the photovoltaic output power distribution instruction, energy storage charge and discharge power instruction, and charging pile power scheduling instruction in the dynamic scheduling policy into executable device control parameters adapted to three core devices in the microgrid, namely photovoltaic inverters, energy storage systems, and charging piles. Specifically, it includes the photovoltaic inverter power output threshold, the energy storage system charge and discharge current limit, and the maximum allowable output power of the charging pile; The charging terminal demand management module is used to collect the electricity usage requests of the charging terminals in the charging and swapping microgrid, obtain the electricity usage request parameters including the charging power demand and the electricity usage priority identifier, and convert the electricity usage request parameters including the charging power demand and the electricity usage priority identifier into load demand signals recognizable by the charging pile orderly control module; The charging pile orderly control module dynamically adjusts the output power and power supply priority of each charging pile in the charging and swapping microgrid according to the maximum allowable output power parameter of the charging pile disassembled by the policy receiving and parsing module, and combines the load demand signal of the charging terminal demand management module to execute orderly charging control; The energy storage system scheduling module adjusts the charge and discharge state and charge and discharge rate of the energy storage system supporting the swapping station in real time according to the energy storage charge and discharge current limit parameter disassembled by the policy receiving and parsing module; The supply-demand matching monitoring module is used to collect the actual photovoltaic output data, the real-time charging and swapping load data, and the energy storage system state data in real time, monitor the matching degree of the photovoltaic output and the charging and swapping load, and provide data feedback for subsequent fine-tuning of control parameters.

[0014] As a further improvement of this technical solution, the process of the photovoltaic accommodation-charging and swapping collaborative control unit to achieve dynamic matching of photovoltaic output and charging and swapping load includes the following steps: S31.1. Receive the dynamic scheduling policy output by the AI-driven dynamic optimization scheduling unit, and disassemble the photovoltaic output power distribution instruction, energy storage charge and discharge power instruction, and charging pile power scheduling instruction in the policy into executable device control parameters adapted to the photovoltaic inverters, energy storage systems, and each charging pile in the microgrid respectively; S31.2. Collect the electricity usage requests of the charging terminals in the charging and swapping microgrid, obtain the electricity usage request parameters including the charging power demand and the electricity usage priority identifier, and convert the electricity usage request parameters including the charging power demand and the electricity usage priority identifier into load demand signals available for charging pile control; S31.3. Combine the maximum allowable output power parameter of the charging pile and the load demand signal, dynamically adjust the output power and power supply priority of each charging pile, and preferentially allocate the power quota corresponding to the photovoltaic output to the charging terminals. The power gap part is supplemented by the energy storage system supporting the swapping station; S31.4 According to the charging and discharging current limit parameters of the energy storage system, when the actual output of photovoltaic power is greater than the charging and swapping load, control the energy storage system of the swapping station to perform charging operation; when the actual output of photovoltaic power is less than the charging and swapping load, control the energy storage system of the swapping station to perform discharging operation to compensate for the power gap. S31.5 Real-time acquisition of actual photovoltaic output data, real-time charging and swapping load data, and status data of the energy storage system supporting the swapping station; monitoring the matching degree between photovoltaic output and charging and swapping load; if the difference between the two exceeds the preset fluctuation range, the deviation data is fed back to the AI-driven dynamic optimization scheduling unit for secondary optimization of the dynamic scheduling strategy.

[0015] As a further improvement to this technical solution, the status monitoring and feedback unit includes a microgrid status acquisition module, a deviation analysis module, and a strategy update triggering module, wherein: The microgrid status acquisition module is used to collect real-time data on the actual operating status of the charging and swapping microgrid, covering core operating parameters of the microgrid including actual photovoltaic output, actual consumption of charging and swapping load, actual charging and discharging power of the energy storage system, and actual output power of the charging stack. The deviation analysis module is used to compare the actual operating status data collected by the microgrid status acquisition module with the predicted value in the dynamic scheduling strategy output by the AI-driven dynamic optimization scheduling unit, calculate the deviation value, and determine whether it exceeds the preset deviation threshold. The strategy update triggering module is used to trigger the AI-driven dynamic optimization scheduling unit to re-calculate and update the scheduling strategy when the deviation analysis module determines that the deviation exceeds the preset deviation threshold, thereby constructing a closed-loop scheduling mechanism.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention utilizes an AI-driven dynamic optimization scheduling unit with a multi-timescale long short-term memory network (constructing a short-term prediction sub-network adapted to real-time matching of charging load and a medium-to-long-term prediction sub-network adapted to energy storage system planning, and dynamically allocating weights using an attention weight fusion mechanism), and an uncertainty quantification module (embedding the Monte Carlo dropout method to perform multiple independent forward propagations, calculate predicted values ​​and variances to generate confidence intervals). Simultaneously, it relies on a data acquisition and preprocessing unit to denoise and standardize data such as photovoltaic output. This effectively solves the problem of existing technologies lacking multi-timescale prediction and uncertainty quantification of photovoltaic output, providing photovoltaic output information tailored to the operational needs of different time periods for microgrid scheduling, and supporting subsequent scheduling strategies to accurately adapt to the dynamic characteristics of charging and swapping microgrids. 2. The present invention effectively solves the problem that the prior art lacks the ability to adaptively adjust the scheduling strategy based on prediction risk through the risk-aware reinforcement learning module (adding the confidence interval width as a state variable for improving the deep deterministic policy gradient algorithm) and the policy adaptive adjustment module (calculating the dynamic risk weight term based on the confidence interval width and introducing a reward function, and preferentially sampling high-uncertainty scheduling experiences from the experience replay pool) in the AI-driven dynamic optimization scheduling unit. The scheduling strategy can dynamically adjust the reward function and exploration rate parameters according to the prediction risk of photovoltaic power output, improving the response flexibility to the fluctuations of photovoltaic power output and the random changes of charging and discharging loads, and reducing the situation of supply-demand imbalance; 3. Through the state monitoring and feedback unit, the present invention can collect core operation parameters such as the actual photovoltaic power output and the actual consumption of charging and discharging loads in real time, compare the deviation between the actual value and the predicted value of the scheduling strategy, and trigger the AI-driven dynamic optimization scheduling unit to re-iterate when the threshold is exceeded, constructing a closed-loop scheduling mechanism. This further solves the problem that the prior art scheduling strategy cannot dynamically iterate according to the actual operation deviation, ensuring the continuous adaptation of the scheduling strategy to the actual operation state of the microgrid, and avoiding the situations of unstable microgrid bus voltage, photovoltaic power abandonment or the inability to meet the charging demand in time caused by prediction deviation, improving the operation stability of the microgrid; <00001​​​​​​​​​​​​​​​​4. Status monitoring and feedback unit; 41. Microgrid status acquisition module; 42. Deviation analysis module; 43. Strategy update triggering module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this embodiment provides an AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids, including: The data acquisition and preprocessing unit 1 collects real-time operating data related to photovoltaic array output, charging and swapping load, microgrid bus electrical parameters and electricity price through distributed sensors, smart meters and microgrid monitoring system. The collected data is denoised and standardized preprocessed to generate a standardized dataset and stored in the time-series real-time database.

[0020] In this embodiment, the data acquisition and preprocessing unit 1 includes a data acquisition module 11 and a data preprocessing and storage module 12, wherein: The data acquisition module 11 collects photovoltaic panel temperature, light intensity and photovoltaic array output power through distributed sensors, collects real-time power, current and electricity price data of charging and swapping loads through smart meters, and collects microgrid bus voltage, current, total active power and reactive power through microgrid monitoring system. Specifically, the data acquisition module 11 acquires data in the following ways: Distributed sensors are deployed in the photovoltaic panel area to collect data on photovoltaic panel temperature, light intensity, and photovoltaic array output power, with a sampling frequency of 1 minute / time. Smart meters are deployed at the inlet of charging and swapping equipment to collect real-time power, current and electricity price data of the charging and swapping load at a frequency of 5 minutes per time. The microgrid monitoring system is deployed in the microgrid distribution room, collecting microgrid bus voltage, current, total active power, and reactive power at a frequency of 10 seconds per acquisition. All collected data is aggregated to the core processing unit of the module via an industrial communication network, achieving preliminary aggregation and format standardization of multi-source data.

[0021] The data preprocessing and storage module 12 uses the 3σ criterion to remove outliers from the various raw data acquired by the data acquisition module 11. It then uses the min-max normalization method to map the denoised data to a unified numerical range, generating a standardized dataset. This standardized dataset is then stored in a time-series real-time database according to the dimension of "data type-acquisition time-device number", supporting the subsequent AI-driven dynamic optimization scheduling unit 2 to quickly query and call the data.

[0022] Specifically, the data preprocessing and storage module 12 uses the 3σ criterion to remove outliers, and the specific process is as follows: For a certain type of raw data (such as the time series of photovoltaic array output power), first calculate its arithmetic mean. and standard deviation ; If a data point satisfies "data point or data points If the value is "outlier", then the data point is considered an outlier and removed.

[0023] Furthermore, min-max normalization is used to map the denoised data to the [0,1] interval, with the specific formula as follows: ,in, The original data values, The minimum value of this type of data. The maximum value of this type of data. These are the normalized data values.

[0024] The AI-driven dynamic optimization scheduling unit 2 is based on a standardized dataset. It constructs a photovoltaic power output prediction model through a multi-timescale long short-term memory network, outputs the photovoltaic power output prediction value and its confidence interval, and uses the confidence interval as the input state variable of the improved deep deterministic strategy gradient algorithm. It introduces a dynamic risk weight term based on the confidence interval width, adaptively adjusts the reward function and exploration rate parameters, and generates a dynamic scheduling strategy that adapts to the dynamic characteristics of the microgrid. It is worth noting that the AI-driven dynamic optimization scheduling unit 2 in this embodiment is deployed at the edge, and its configuration example is as follows: The NVIDIA Jetson AGXXavier edge computing device is equipped with a Carmel architecture ARMv8 264-bit 8-core CPU (with a maximum single-core frequency of 2.26GHz), which can meet the real-time requirements of data preprocessing and scheduling strategy logic operations. Equipped with a 512-core NVIDIA Volta architecture GPU (integrated with TensorCores, supporting FP16 and INT8 precision AI acceleration computing), it can provide 32 trillion calculations per second (32 TFLOPS) of computing power for the training and inference of the LSTM photovoltaic power output prediction model and the improved DDPG scheduling algorithm. Equipped with 32GB LPDDR4x high-speed memory (137GB / s bandwidth) to ensure parallel access to data and model parameters across multiple time scales. For storage, it is recommended to configure at least 128GB eMMC embedded storage or an external 512GB NVMeSSD (for storing time-series databases, AI model weight files, and system logs). This hardware configuration can fully meet the real-time computing requirements of "multi-timescale LSTM prediction + improved DDPG scheduling" in this system, and realize millisecond-level policy updates and device control command issuance at the edge.

[0025] In this embodiment, the AI-driven dynamic optimization scheduling unit 2 includes a multi-timescale prediction module 21, an uncertainty quantification module 22, a risk perception reinforcement learning module 23, and a policy adaptive adjustment module 24, wherein: The multi-timescale prediction module 21 uses a multi-time-window long short-term memory network to realize short-term and medium-to-long-term prediction of photovoltaic power output and outputs the prediction confidence interval. The uncertainty quantification module 22 calculates the photovoltaic output prediction confidence interval based on the prediction feature sequence; The risk perception reinforcement learning module 23 introduces confidence interval information as an input state variable into the improved deep deterministic policy gradient algorithm; The policy adaptive adjustment module 24 dynamically adjusts the reward function weights and exploration rate parameters of the reinforcement learning algorithm based on the confidence interval, so as to realize the generation of dynamic scheduling strategy.

[0026] In this embodiment, the process of the multi-timescale prediction module 21 outputting the prediction confidence interval includes the following steps: S21.1 Set short-term and medium-to-long-term time windows, and construct short-term prediction sub-networks and medium-to-long-term prediction sub-networks respectively. Both short-term and medium-to-long-term prediction sub-networks adopt the long short-term memory network structure. S21.2 Input short-term photovoltaic power output historical data from the standardized dataset into the short-term prediction subnetwork, and input medium- and long-term photovoltaic power output historical data and corresponding meteorological data from the standardized dataset into the medium- and long-term prediction subnetwork. Minimize the error between the predicted value and the actual value through the gradient descent method, and iterate until the error of the validation set converges and stabilizes. S21.3. The prediction outputs of the short-term prediction sub-network and the medium- and long-term prediction sub-network are weighted and fused through the attention weight fusion mechanism. During the fusion, the weight ratio of the outputs of the short-term prediction sub-network and the medium- and long-term prediction sub-network is dynamically allocated according to the microgrid's requirements for prediction accuracy at different time periods to obtain the preliminary photovoltaic power output prediction value. S21.4 Transmit the preliminary photovoltaic power output prediction value after fusion to the uncertainty quantification module 22.

[0027] Specifically, to accurately match the dual needs of "real-time adjustment of short-term charging load and planning of medium- and long-term energy storage systems" in photovoltaic-integrated charging and swapping microgrids, the multi-time-scale prediction module 21 adopts a "dual LSTM sub-network + attention weight fusion" architecture to construct a photovoltaic output prediction model. Details are as follows: To meet the short-term forecasting needs within 1 hour (adapting to real-time power allocation of the charging pile, requiring minute-level accuracy), a 2-layer LSTM short-term forecasting sub-network is constructed, with an input sequence length of 60 (corresponding to the historical photovoltaic power output data collected by data acquisition module 11 once every 1 minute for a total of 60 minutes). To meet the medium- and long-term forecasting needs within 24 hours (adapting to the energy storage system's charge and discharge plan formulation, requiring hourly trend control), a 3-layer LSTM medium- and long-term forecasting sub-network is constructed. The input sequence length is 1440 (corresponding to the historical photovoltaic power output data collected by data acquisition module 11 every 5 minutes for a total of 24 hours, and simultaneously integrating hourly meteorological data for that period—including average light intensity and average ambient temperature—to improve the accuracy of medium- and long-term trend forecasting). Meanwhile, both the short-term and medium-to-long-term prediction subnetworks employ the Adam optimizer, with an initial learning rate of 0.001 and mean squared error as the loss function. During iterative training, the training is performed based on the validation set loss fluctuating less than 0.001 for five consecutive rounds. "As a convergence criterion, it ensures that the subnetwork has stable predictive capabilities."

[0028] Furthermore, during the prediction phase, the outputs of the two sub-networks are integrated through an attention weight fusion mechanism: the prediction result of the short-term sub-network is used as the query vector. The prediction results of the medium- and long-term subnetworks are also used as key vectors. Sum value vector Attention formula by scaling the dot product The similarity between the two is calculated, and then the similarity is normalized into weights using the Softmax function (ensuring that the sum of the weights is 1), finally yielding the preliminary photovoltaic power output prediction value after fusion; where The key vector dimension is set to 256 to avoid gradient vanishing due to excessively large dot product results. For example, during peak charging load periods at midday, the microgrid requires higher short-term prediction accuracy, and the attention mechanism automatically increases the weight ratio of the short-term sub-network output; during nighttime energy storage planning periods, the weight ratio of the medium- and long-term sub-network output is increased, achieving dynamic adaptation of "allocating prediction weights on demand".

[0029] In photovoltaic charging and swapping microgrid scheduling, traditional technologies either employ single-time-scale prediction models (such as focusing only on minute-level or hour-level photovoltaic output) or fuse multi-scale results with fixed weights. This makes it difficult to adapt to the dynamic demands of "precise matching of charging load in the short term and support for energy storage planning in the medium to long term." Relying on medium- to long-term predictions during peak charging times can lead to charging gaps due to insufficient accuracy, while relying on short-term predictions for energy storage planning can result in incomplete planning due to limited coverage. This embodiment employs a "dual LSTM sub-network (short-term focusing on minute-level accuracy to adapt to real-time power allocation of charging piles, and medium- to long-term focusing on hourly trends to support energy storage charging and discharging plans) + scaling dot product attention fusion" approach. This allows the two sub-networks to focus on scheduling needs within different time windows. Furthermore, through an attention mechanism, it acts like an "intelligent dispatcher," automatically adjusting the output weight ratio of the two sub-networks based on the microgrid's operating scenario (such as peak charging times and energy storage planning periods). Its core function is to increase the weight of short-term sub-networks during the midday charging peak to ensure the prediction accuracy of real-time power supply of the charging pile; and to increase the weight of medium- and long-term sub-networks during nighttime energy storage planning to provide a reference for the energy storage system's output trend throughout the day. At the same time, the independent training and dynamic fusion of the two sub-networks also reduces the risk of prediction collapse caused by sudden weather (such as cloudy days) in a single model, improves the robustness of the prediction system, and provides accurate prediction support for the scheduling of photovoltaic-based charging and swapping microgrids at different times.

[0030] In this embodiment, the process by which the uncertainty quantification module 22 calculates the confidence interval for photovoltaic power output prediction includes the following steps: S22.1. The Monte Carlo dropout method is embedded in the long short-term memory network of the multi-timescale prediction module 21 to ensure that the dropout layer in the network training phase continues to be effective in the prediction phase. S22.2 The input data of the multi-timescale prediction module 21 is processed through a long short-term memory network that embeds the Monte Carlo dropout method and keeps the dropout layer active for multiple independent forward propagations to obtain multiple parallel prediction output values. S22.3 Statistical analysis is performed on the output values ​​of multiple parallel predictions to calculate the final photovoltaic power output prediction value. and variance ; S22.4, combined with confidence coefficient (Selected based on the operational stability requirements of microgrid equipment), construct the confidence interval for photovoltaic power output prediction, and calculate the width of the confidence interval. Confidence interval and width Output to Risk Perception Reinforcement Learning Module 23.

[0031] Specifically, to address the scheduling risk caused by "providing only numerical values ​​without reliability in the prediction results," the uncertainty quantification module 22 uses the Monte Carlo dropout method to quantify the uncertainty of the initial prediction values ​​output by the multi-timescale prediction module 21. The details are as follows: First, a dropout layer is embedded in the hidden layer of the LSTM network in the multi-timescale prediction module 21, with a dropout rate of 0.2 (a conventional value that balances "uncertainty quantization accuracy" and "computational efficiency," avoiding both insufficient quantization due to an excessively low dropout rate and divergent prediction values ​​due to an excessively high dropout rate). Special attention should be paid to the following: during the model training phase, the dropout layer randomly drops some neurons as usual; during the prediction phase, the dropout layer must remain active (existing Monte Carlo dropout methods often neglect to enable the dropout layer during the prediction phase, leading to quantization failure), ensuring that the network structure has random differences in each prediction, thereby simulating prediction uncertainty. Then, for the same set of input data (i.e., the standardized input data of the multi-timescale prediction module 21), 100 independent forward propagations are performed through the LSTM network with the embedded dropout layer (100 times is a value that balances "statistical accuracy" and "computational time"; too few times will result in inaccurate variance calculation, while too many times will lead to microgrid scheduling delay), resulting in 100 different parallel prediction values. ; A statistical analysis was performed on the above 100 predicted values: the arithmetic mean was calculated, and this arithmetic mean was used as the final predicted photovoltaic output value. The formula is ; Calculate sample variance The formula is (The denominator is set to 99 for an unbiased estimate); where, Indicates the first The predicted value of the second forward propagation; Next, confidence coefficients are selected based on the operational stability requirements of microgrid devices. (This embodiment selects) (Corresponding to a 95% confidence level, meaning there is a 95% probability that the actual photovoltaic output falls within this range), construct the confidence interval. And calculate the interval width. (The wider the width, the higher the prediction uncertainty), the formula is: ; Finally, the "final predicted value" Confidence interval, interval width "This data is also output to the risk perception reinforcement learning module 23, providing a quantitative basis for the 'risk perception' of subsequent scheduling strategies. For example, when..." When the value is small, the prediction uncertainty is low, and scheduling can rely more on the prediction value; when When the value is large, the forecast uncertainty is high, and more energy storage backup capacity needs to be reserved for dispatch.

[0032] Traditional technologies only output "point prediction values" of photovoltaic (PV) output, failing to explain the reliability of the prediction results or provide dispatchers with a basis for risk buffering. When the predicted value deviates significantly from the actual output, it can easily lead to charging interruptions or PV curtailment, leaving dispatchers to rely on experience to judge the accuracy of the prediction. This embodiment employs a "Monte Carlo dropout method (embedding a dropout layer in the LSTM hidden layer and maintaining it throughout the prediction phase) + 100 independent forward propagations + statistical analysis" approach. After performing multiple independent predictions on the same input, the final predicted value, variance, confidence interval, and interval width are calculated, essentially equipping the prediction results with a "risk detector." Its core function is to quantify the abstract "prediction uncertainty" into a calculable numerical indicator. When encountering cloudy weather, the "detector" will provide a larger interval width, indicating to the dispatcher that "the current prediction risk is high, and more reserve capacity should be allocated to energy storage." On sunny days, the interval width is smaller, allowing the dispatcher to confidently prioritize PV output allocation to charging piles. This quantification transforms microgrid scheduling from "blindly relying on forecasts" to "making decisions with risk awareness," reducing supply and demand imbalances during high-risk periods and increasing photovoltaic absorption rates during low-risk periods.

[0033] Traditional technologies only output "point prediction values" of photovoltaic (PV) output, failing to explain the reliability of the prediction results or provide dispatchers with a basis for risk buffering. When the predicted value deviates significantly from the actual output, it can easily lead to charging interruptions or PV curtailment, leaving dispatchers to rely on experience to judge the accuracy of the prediction. This embodiment employs a "Monte Carlo dropout method (embedding a dropout layer in the LSTM hidden layer and maintaining it throughout the prediction phase) + 100 independent forward propagations + statistical analysis" approach. After performing multiple independent predictions on the same input, the final predicted value, variance, confidence interval, and interval width are calculated, essentially equipping the prediction results with a "risk detector." Its core function is to quantify the abstract "prediction uncertainty" into a calculable numerical indicator. When encountering cloudy weather, the "detector" will provide a larger interval width, indicating to the dispatcher that "the current prediction risk is high, and more reserve capacity should be allocated to energy storage." On sunny days, the interval width is smaller, allowing the dispatcher to confidently prioritize PV output allocation to charging piles. This quantification transforms microgrid scheduling from "blindly relying on forecasts" to "making decisions with risk awareness," reducing supply and demand imbalances during high-risk periods and increasing photovoltaic absorption rates during low-risk periods.

[0034] In this embodiment, the process of embedding confidence interval information in the risk perception reinforcement learning module 23 includes the following steps: S23.1. Based on the original state variables of the improved deep deterministic policy gradient algorithm, a new confidence interval width is added. As state variables, they expand the state input dimension of the algorithm; S23.2, Based on the width of the confidence interval Width of the historical maximum confidence interval The ratio is used to classify uncertainty levels and set the exploration rate. The range of values, where a first threshold is set. Second threshold ,and : when hour, Take the lower range of values; when hour, Take the middle range of values; when hour, Take the higher range of values; S23.3. Input the expanded complete state variables into the actor network of the algorithm, so that the reinforcement learning agent can directly perceive the uncertainty of prediction during the decision-making process.

[0035] The core objective of the risk-aware reinforcement learning module 23 is to enable the reinforcement learning algorithm to "aware of the risks in photovoltaic power output prediction and dynamically adjust the microgrid scheduling strategy." To this end, two key improvements are made to the classic Deep Deterministic Policy Gradient (DDPG) algorithm: First, the confidence interval width is increased. The two approaches are: first, incorporating the state space to quantify and predict risks; and second, dynamically adjusting the exploration rate based on the risk level to adapt to microgrid scenarios.

[0036] Specifically, the state space of the classic DDPG only includes "predicted photovoltaic output, charging / swapping load, and energy storage state of charge," without considering the uncertainty of the prediction results. This improvement increases the confidence interval width. Introduced as a new state variable, this allows the reinforcement learning agent to directly "perceive" and predict risks during decision-making. The expanded state space is as follows: ; in, represent Forecasted charging and swapping load at any time (unit: kW, from data acquisition and preprocessing unit 1). represent State of charge of energy storage system at any time (unit: %) represent The width of the confidence interval for the predicted photovoltaic output at any given time (unit: kW, from uncertainty quantification module 22), and The wider the width, the higher the uncertainty of the prediction; Specifically, DDPG is based on an "Actor-Critic" framework. The actor network outputs continuously scheduled actions, and the critic network evaluates the value of the actions. The actor network uses a 3-layer fully connected neural network, with the expanded state as input. (4-dimensional vector), the output is a continuous action vector. (Such as photovoltaic power output allocation ratio, energy storage charging and discharging power, charging pile priority, etc.), the actions are determined by the strategy function. The formula for generation is: ; in, , , This represents the linear transformation of the fully connected layer (including the weight matrix and bias terms). Parameters representing the actor network, This represents the activation function, used to map action values ​​to a reasonable range (such as photovoltaic allocation ratio [0,1], energy storage power [-50,50]kW). Specifically, the critic network also uses a 3-layer fully connected neural network, with the input being the state. and actions The output is a state-action value. The formula is: ; in, Linear transformation layers representing states and actions, respectively. This represents the concatenation operation between state and action vectors. This represents a linear transformation of the subsequent fully connected layer. The parameters represent the network of critics.

[0037] Furthermore, addressing the shortcomings of classic DDPG, which employs a fixed exploration strategy (such as adding Ornstein-Uhlenbeck noise) and cannot adapt to dynamic changes in predicted risk, this improvement is based on... Width of the historical maximum confidence interval The ratio of the exploration rate is dynamically adjusted. The formula is: ; in, Represents the exploration rate (unitless), used to control the ratio of "exploring new strategies" to "utilizing known strategies"; , , These represent the exploration rate ranges for low, medium, and high-risk scenarios, respectively (in this example, we take...). ); , The threshold representing the risk level (in this embodiment, it is taken as...) (Based on statistics from historical microgrid operation data) The maximum confidence interval width since operation began (unit: kW) serves as the benchmark for risk level. In implementation, Directly affects the intensity of motion noise: The larger the value, the stronger the added noise, prompting the agent to explore more potential strategies; The smaller the value, the weaker the noise, and the more likely the agent is to utilize proven and reliable strategies.

[0038] Furthermore, the gradient updates for the actor network and the critic network are consistent with the classic DDPG, with core formulas including the critic network loss function and the actor network policy gradient. The critic network loss function (mean squared error) is: ; The gradient of the agent network policy is: ; In the above formula, Represents the experience replay pool (storage) (tuples) represent The reward value at each moment (from the policy adaptive adjustment module 24) ), Representative discount factor (in this example, it is taken as) =0.99)), These represent the outputs of the target commentator network and the target actor network, respectively. These represent the parameters of the target network (synchronized with the current network via soft updates).

[0039] In traditional reinforcement learning scheduling, the algorithm is unaware of predicted risks, and the exploration rate remains constant. Decisions based on a fixed exploration rate during high-risk scenarios can easily lead to failures due to overly aggressive strategies; while blind exploration during low-risk scenarios can cause frequent scheduling fluctuations (such as frequent adjustments to the charging pile power). This embodiment employs an "extended state space (adding confidence interval width)." By using the method of "as a state variable + dynamically adjusting the exploration rate," the reinforcement learning agent can directly obtain information about "prediction uncertainty" when making decisions, and then... The risk level is determined by the ratio to the historical maximum width, and the exploration rate is dynamically adjusted, much like equipping the algorithm with a "risk radar." Its core function is to increase the exploration rate in high-risk scenarios (such as large fluctuations in sunlight before a typhoon), actively exploring conservative strategies like "more backup power and more reliance on the grid"; and to decrease the exploration rate in low-risk scenarios (such as stable sunlight at midday on a sunny day), prioritizing mature strategies like "more photovoltaic power allocation and more energy storage charging." This design transforms the reinforcement learning agent from "ignorant and fearless" to "aware of risks and adept at adjustment," significantly reducing the probability of charging interruptions while improving photovoltaic power absorption rates.

[0040] In this step, the process of the strategy adaptive adjustment module 24 optimizing algorithm parameters and generating a scheduling strategy includes the following steps: S24.1, Confidence interval width based on the output of uncertainty quantization module 22 Width of the historical maximum confidence interval Calculate dynamic risk weights ; S24.2, Dynamic risk weights Introducing a reward function to form a multi-objective technical optimization reward function ; S24.3. Introduce a priority sampling rule and set a threshold for the empirical replay pool of the improved deep deterministic strategy gradient algorithm. ,right Scheduling experience is assigned a higher sampling weight, and the actor-evaluator network parameters of this type of experience-based update algorithm are extracted first; among which... satisfy ; S24.4 Based on the output of the optimized and improved deep deterministic strategy gradient algorithm, combined with the rated power of the microgrid photovoltaic inverter, the charge and discharge rate limit of the energy storage system, and the capacity constraint of the charging pile, a dynamic scheduling strategy including photovoltaic power distribution instructions, energy storage charge and discharge power instructions, and charging pile power scheduling instructions is generated and output to the photovoltaic consumption-charging and swapping collaborative control unit 3.

[0041] The core objective of the strategy adaptive adjustment module 24 is to enable the scheduling strategy to be "dynamically optimized according to the predicted risk of photovoltaic output". By dynamically adjusting the reward function with risk weight and prioritizing high-risk experience, a scheduling strategy adapted to the dynamic characteristics of the microgrid is finally generated.

[0042] Specifically, in order to balance the scheduling objectives of "photovoltaic consumption" and "load guarantee" under different risk scenarios, the strategy adaptive adjustment module 24 first calculates the dynamic risk weight. The formula is: ; in, Represents dynamic risk weights (unitless), with values ​​ranging from... , The larger The larger the value, the higher the current predicted risk; Based on the above calculations Construct a multi-objective optimization reward function The formula is: ; in, express The multi-objective reward value at any given moment (unitless), the larger the value, the better the current scheduling strategy; express Photovoltaic grid integration rate at all times ,and ; express Real-time load shortage rate ,and ; This represents the basic weighting coefficient (in this embodiment, we take...). (This is a unitless term used to balance the goals of "photovoltaic consumption" and "load power supply guarantee"); when When it is high-risk, through The weight of "reducing the load shortage rate" is increased (prioritizing charging demand, even at the cost of sacrificing some photovoltaic power consumption); when When the risk is low, this weight is reduced, and priority is given to increasing the photovoltaic absorption rate.

[0043] Specifically, to ensure that reinforcement learning algorithms "prioritize learning from high-risk scenario experiences," this module introduces a priority sampling rule for the experience replay pool, based on... and Different sampling priorities are assigned based on the relationship, as shown in the formula: ; ; in, express The sampling probability (unitless) of the time-scheduling experience satisfies ; express The time-series difference error at any given moment is dimensionless and reflects the "updating value" of experience. ,in The state value output by the critic network. Discount factor; Represents a small constant (in this embodiment) ,avoid The sampling probability is 0. Indicates the priority coefficient (in this embodiment) ), control the degree of influence of priority, It degenerates into uniform sampling; Indicates the priority sampling threshold (in this embodiment, it is taken as...). This is used to distinguish between "high-risk experience" and "low-risk experience," with high-risk experience being assigned a higher sampling weight. Represents the total amount of experience in the experience replay pool (unitless); The above design ensures that the sampling probability of high-risk experiences is 2 to 3 times higher than that of low-risk experiences, thus ensuring that the algorithm prioritizes learning "how to deal with high-risk scenarios" and avoids high-risk experiences failing to update the network due to low sampling probability.

[0044] Furthermore, considering the security constraints of microgrid devices, the "action vector" output by the improved DDPG algorithm is decomposed into executable device control commands: Photovoltaic power output allocation command: Based on the photovoltaic power output allocation ratio output by the algorithm. Combined with the rated power of the photovoltaic inverter Calculate the power distribution according to the following formula. ; ; Energy storage charging and discharging power command: based on the energy storage power output by the algorithm. Combined with energy storage charge / discharge rate With energy storage rated capacity ,pass Function limits power range ; Charge pile power scheduling instruction: Based on the power of the i-th charge pile output by the algorithm Combined with the charging pile capacity Calculate the actual power using the formula; ; The above instructions are integrated into a dynamic scheduling strategy and output to the photovoltaic consumption-charging and swapping collaborative control unit 3 to achieve a closed loop of "risk perception-strategy optimization-equipment execution".

[0045] In traditional reinforcement learning, the target weights of the reward function are fixed, and experience replay does not consider the importance of the scenario—in high-risk scenarios, focusing solely on "improving the grid connection rate" can lead to charging interruptions, while in low-risk scenarios, conservative strategies waste photovoltaic output. Furthermore, high-risk experiences are overwhelmed by a large number of low-risk experiences, preventing the algorithm from learning methods to handle high-risk situations. This embodiment uses a "dynamic risk weight adjustment reward function (based on...") Calculate risk weight adjustment reward target priority) + priority sampling experience (for By means of "assigning a higher sampling probability to high - risk experiences with a proportion exceeding 0.5)", it is like hiring an "intelligent coach" for the algorithm. Its core function is that when in a high - risk situation (such as rainy weather), the reward function gives priority to ensuring charging, increases the grid energy supplement and energy storage backup, and significantly reduces the charging interruption rate; when in a low - risk situation (such as sunny days), the reward function gives priority to consuming photovoltaic power, reduces the grid energy supplement, increases the energy storage charging, and improves the photovoltaic power consumption rate. At the same time, the "coach" focuses on tutoring high - risk experiences, enabling the algorithm to master strategies for dealing with high - risks faster, reducing the scheduling error in high - risk scenarios, and making the microgrid both reliable and efficient.

[0046] The photovoltaic power consumption - charging and swapping协同控制单元3 (Photovoltaic power consumption - charging and swapping coordinated control unit 3), the photovoltaic power consumption - charging and swapping coordinated control unit 3 receives the dynamic scheduling strategy, performs orderly charging control on the charging piles, and performs energy storage scheduling on the energy storage system supporting the swapping station, to achieve the dynamic matching of photovoltaic output and charging and swapping loads; In this embodiment, the photovoltaic power consumption - charging and swapping coordinated control unit 3 includes a strategy receiving and parsing module 31, a charging terminal demand management module 32, a charging pile orderly control module 33, an energy storage system scheduling module 34, and a supply - demand matching monitoring module 35, where: The strategy receiving and parsing module 31 is used to receive the dynamic scheduling strategy output by the strategy adaptive adjustment module 24, and disassemble the photovoltaic output allocation instruction, the energy storage charge - discharge power instruction, and the charging pile power scheduling instruction in the dynamic scheduling strategy into executable device control parameters suitable for three core devices in the microgrid, namely photovoltaic inverters, energy storage systems, and charging piles, specifically including the photovoltaic inverter power output threshold, the energy storage system charge - discharge current limit, and the maximum allowable output power of the charging pile; Specifically, the strategy receiving and parsing module 31, as the "instruction center" of unit 3, is responsible for receiving and disassembling the dynamic scheduling strategy, as follows: Strategy receiving: Adopting the industrial Ethernet (TCP / IP protocol) communication method, through an industrial - grade PLC configured with an Ethernet interface, receive the dynamic scheduling strategy (transmission format is JSON) from the AI - driven dynamic optimization scheduling unit 2, and set the communication baud rate to 100 Mbps to ensure real - time transmission of instructions.

[0047] Instruction disassembling: Photovoltaic inverter power output threshold: According to the "photovoltaic output allocation instruction" in the dynamic scheduling strategy, combined with the rated power of the photovoltaic inverter, calculate the photovoltaic inverter power output threshold (i.e., the maximum power allowed for the photovoltaic inverter to output), and convert it into a 4 - 20 mA analog quantity instruction recognizable by the photovoltaic inverter (corresponding to a power range of 0 - threshold power).

[0048] Energy storage system charge and discharge current limit: Based on the "energy storage charge and discharge power command", combined with the rated capacity and charge and discharge rate of the energy storage system, the charge and discharge current limit of the energy storage system is calculated, and the current limit command is sent to the energy storage converter through the RS485 interface.

[0049] Maximum allowable output power of charging pile: The value of the "charging pile power scheduling instruction" in the dynamic scheduling strategy is directly extracted and used as the maximum allowable output power of each charging pile. It is sent to the charging pile control unit (the built-in core control component of the charging pile) in digital form (Modbus register value).

[0050] The charging terminal demand management module 32 is used to collect the power consumption requests of charging terminals in the charging and swapping microgrid, obtain power consumption request parameters including charging power demand and power consumption priority identifier, and convert the power consumption request parameters including charging power demand and power consumption priority identifier into load demand signals that can be recognized by the charging pile orderly control module 33. Specifically, the charging terminal demand management module 32 is responsible for collecting and processing the power consumption requests of the charging terminals, as follows: Power demand request collection: The power demand of the charging terminal is collected at a frequency of 5 minutes / time via the CAN bus or RS485 interface of the charging terminal. The collected parameters include charging power demand (e.g., 0-60kW range) and power priority identifier ("emergency" or "normal").

[0051] Load demand signal conversion: The collected power demand parameters are converted into load demand signals that can be recognized by the charging pile orderly control module 33, including the total charging power demand, peak value and number of charging terminals with different priorities, forming a two-dimensional array and priority distribution table, which are encapsulated into Modbus TCP protocol data packets and sent to the charging pile orderly control module 33.

[0052] The charging pile orderly control module 33 dynamically adjusts the output power and power supply priority of each charging pile in the charging and swapping microgrid based on the maximum allowable output power parameters of the charging pile obtained by the strategy receiving and parsing module 31 and the load demand signal of the charging terminal demand management module 32, and executes orderly charging control. Specifically, the charging pile orderly control module 33 dynamically adjusts the charging pile output power and priority based on commands and load requirements, as follows: Power and priority adjustment logic: Extract the maximum allowable output power of the charging pile from the strategy receiving and parsing module 31 and the load demand signal from the charging terminal demand management module 32; If the total charging demand is less than or equal to the maximum allowable output power, the required power will be allocated to emergency charging terminals according to the principle of "emergency priority and power allocation on demand", and the remaining power will be allocated to ordinary charging terminals. If the total charging demand > the maximum allowable output power, first allocate the corresponding proportion of power to the emergency charging terminal, and the remaining power is allocated to the ordinary charging terminals. The gap is marked as "to be supplemented by energy storage".

[0053] Power distribution execution: Through the power regulation interface of the charging stack control unit, send the calculated power value to the charging stack, and at the same time update the power supply priority identifier of the charging stack.

[0054] The energy storage system scheduling module 34 receives and analyzes the energy storage charge and discharge current limit parameters disassembled by the policy receiving and parsing module 31, and real-time regulates the charge and discharge status and charge and discharge rate of the energy storage system supporting the swapping station; Specifically, the energy storage system scheduling module 34 is responsible for regulating the charge and discharge status of the energy storage system supporting the swapping station, as follows: Charge and discharge status judgment: Real-time collect the actual photovoltaic output (from the power feedback of the photovoltaic inverter) and the charge and discharge load (from the power feedback of the total load meter). If the actual photovoltaic output > the charge and discharge load, it is determined as the "charging state"; if the actual photovoltaic output < the charge and discharge load, it is determined as the "discharging state".

[0055] Charge and discharge rate regulation: During charging, according to the energy storage charge and discharge current limit of the policy receiving and parsing module 31, send a charging current command to the energy storage converter to control the energy storage system to charge with this current; During discharging, according to the power gap and the energy storage charge and discharge current limit, send a discharging current command to the energy storage converter to control the energy storage system to discharge to compensate for the power gap.

[0056] The supply-demand matching monitoring module 35 is used to real-time collect the actual photovoltaic output data, the real-time charge and discharge load data and the energy storage system status data, monitor the matching degree between the photovoltaic output and the charge and discharge load, and provide data feedback for subsequent fine-tuning of control parameters.

[0057] Specifically, the supply-demand matching monitoring module 35 is responsible for real-time monitoring and feedback of the supply-demand matching degree, as follows: Data collection: Actual photovoltaic output: Collected through the Ethernet interface of the photovoltaic inverter at a frequency of 10 seconds / time; Charge and discharge load: Collected through the RS485 interface of the total load meter at a frequency of 10 seconds / time; Energy storage system status: Collect data such as SOC and charge and discharge current through the CAN interface of the energy storage BMS at a frequency of 10 seconds / time.

[0058] Matching degree monitoring and feedback: Calculate the difference ratio of the photovoltaic output and the charge and discharge load: ; where It represents the percentage deviation of supply and demand matching, used to quantify the degree of difference between the actual output of photovoltaic power and the actual load of charging and swapping. The larger the value, the lower the matching degree between the two. This indicates the actual output of the photovoltaic system, expressed in "kW". It refers to the actual power output of the photovoltaic inverter, which is collected in real time through the Ethernet interface of the photovoltaic inverter. This indicates the actual load of the charging and swapping system, in units of "kW". It refers to the total power consumption of all charging terminals, auxiliary equipment, etc. in the charging and swapping microgrid, which is collected in real time through the RS485 interface of the load meter. like (Preset fluctuation range) The deviation data (actual photovoltaic output, charging and swapping load, power difference) are encapsulated into MQTT messages and sent to the AI-driven dynamic optimization scheduling unit 2 for secondary optimization of the dynamic scheduling strategy.

[0059] In this embodiment, the process by which the photovoltaic absorption-charging / swapping coordinated control unit 3 achieves dynamic matching between photovoltaic output and charging / swapping load includes the following steps: S31.1 Receive the dynamic scheduling strategy output by the AI-driven dynamic optimization scheduling unit 2, and decompose the photovoltaic power distribution instruction, energy storage charging and discharging power instruction, and charging pile power scheduling instruction in the strategy into executable equipment control parameters that are adapted to the photovoltaic inverter, energy storage system and each charging pile in the microgrid. S31.2 Collect the power consumption requests of charging terminals in the charging and swapping microgrid, obtain power consumption request parameters including charging power demand and power consumption priority identifier, and convert the power consumption request parameters including charging power demand and power consumption priority identifier into load demand signals that can be used for charging pile control. S31.3. Combining the maximum allowable output power parameters of the charging pile with the load demand signal, dynamically adjust the output power and power supply priority of each charging pile, prioritize the allocation of the photovoltaic power quota to the charging terminal, and supplement the power gap through the energy storage system of the battery swapping station. S31.4 According to the charging and discharging current limit parameters of the energy storage system, when the actual output of photovoltaic power is greater than the charging and swapping load, control the energy storage system of the swapping station to perform charging operation; when the actual output of photovoltaic power is less than the charging and swapping load, control the energy storage system of the swapping station to perform discharging operation to compensate for the power gap. S31.5 Real-time acquisition of actual photovoltaic output data, real-time charging and swapping load data, and status data of the energy storage system supporting the swapping station; monitoring the matching degree between photovoltaic output and charging and swapping load; if the difference between the two exceeds the preset fluctuation range, the deviation data is fed back to the AI-driven dynamic optimization scheduling unit 2 for secondary optimization of the dynamic scheduling strategy.

[0060] The status monitoring and feedback unit 4 collects the actual operating status data of the microgrid in real time and compares the deviation with the predicted value in the dynamic scheduling strategy. When the deviation exceeds the preset deviation threshold, the AI-driven dynamic optimization scheduling unit 2 is triggered to re-iterate and update the scheduling strategy, forming a closed-loop scheduling mechanism.

[0061] In this embodiment, the status monitoring and feedback unit 4 includes a microgrid status acquisition module 41, a deviation analysis module 42, and a strategy update triggering module 43, wherein: The microgrid status acquisition module 41 is used to collect the actual operating status data of the charging and swapping microgrid in real time, covering the core operating parameters of the microgrid, including the actual output of photovoltaic power, the actual consumption of charging and swapping load, the actual charging and discharging power of the energy storage system, and the actual output power of the charging pile. Specifically, the microgrid status acquisition module 41, as the data input core of the status monitoring and feedback unit 4, adopts a "multi-interface adaptation + high-frequency acquisition + data preprocessing" approach to reliably acquire the core operating parameters of the microgrid, as detailed below: To assess the actual output power of the photovoltaic system, real-time output power data of the inverter is collected every 10 seconds via the standard Ethernet interface (supporting Modbus TCP protocol) of the photovoltaic inverter. The parameters are identified as follows. The unit is uniformly "kW"; To assess the actual power consumption of the charging and swapping load, the total power consumption is synchronously collected every 10 seconds via the RS485 interface of the smart meter (supporting the DL / T645-2007 protocol) at the main inlet of the microgrid. The parameters are identified as follows. ; To determine the actual charging and discharging power of the energy storage system, the real-time power signal of the energy storage converter is collected via the CAN bus interface of the battery management system (BMS) (charging is considered positive and discharging is considered negative). The parameters are identified as follows: ; For the actual output power of the charging pile, the real-time output power of each charging pile is collected through the Ethernet interface (compatible with Modbus TCP protocol) of each charging pile control unit and then summarized into the total power. The parameter is identified as follows. .

[0062] Furthermore, all collected data is stored in a local industrial-grade database (such as SQLite) in the format of "timestamp + parameter type + value + unit". For null values ​​or abrupt outliers caused by communication interruptions, linear interpolation is used to complete the data, ensuring the continuity and validity of the data and providing reliable input for subsequent deviation analysis.

[0063] The deviation analysis module 42 is used to compare the actual operating status data collected by the microgrid status acquisition module 41 with the predicted value in the dynamic scheduling strategy output by the AI-driven dynamic optimization scheduling unit 2, calculate the deviation value, and determine whether it exceeds the preset deviation threshold. Specifically, the deviation analysis module 42 uses "classification calculation of deviation + threshold grading judgment" as its core logic to achieve accurate comparison between the actual operating status and the predicted value of the scheduling strategy, as follows: First, from the policy output interface of the AI-driven dynamic optimization scheduling unit 2, the predicted values ​​of the corresponding parameters in the dynamic scheduling policy are synchronously obtained (denoted as follows): ); Then, for each type of parameter, the degree of deviation is calculated using the relative deviation formula, which is as follows: ,in For parameters The relative deviation (unit: %). These are the actual collected values. This is the predicted value for the strategy; Furthermore, to adapt to the operating characteristics of different parameters, differentiated deviation thresholds are set based on microgrid equipment debugging experience: Photovoltaic power output fluctuates significantly due to the influence of natural sunlight; the threshold is set to... The charging and battery swapping load changes dynamically with user demand; the threshold is set to... ; The charging and discharging power of the energy storage system is limited by the SOC, resulting in a regulation delay. The threshold is set to... ; The charging pile has high power response accuracy, and the threshold is set to .

[0064] Furthermore, the judgment logic adopts "OR logic", that is, when the relative deviation of any type of parameter exceeds the corresponding threshold, the analysis result of "deviation exceeds limit" is immediately generated, and the deviation parameter type, specific deviation value and current timestamp are encapsulated into an analysis report and transmitted to the strategy update trigger module 43.

[0065] The strategy update triggering module 43 is used to trigger the AI-driven dynamic optimization scheduling unit 2 to re-iterate and update the scheduling strategy when the deviation analysis module 42 determines that the deviation exceeds the preset deviation threshold, thereby constructing a closed-loop scheduling mechanism.

[0066] Specifically, the strategy update trigger module 43 is the core of the execution of the closed-loop scheduling mechanism, designed based on the principle of "fast response + precise triggering", as follows: The strategy update trigger module 43 has a built-in communication interface that matches the AI-driven dynamic optimization scheduling unit 2, and uses the industrial-grade MQTT protocol (message service quality level QoS=1) to ensure the reliability of signal transmission. Upon receiving the "deviation exceeding limits" analysis report from the deviation analysis module 42, a standardized "strategy update trigger signal" is immediately generated. The signal content includes the deviation parameter identifier, the actual deviation percentage, a snapshot of the current microgrid core operating parameters, and a trigger timestamp, ensuring that the AI-driven dynamic optimization scheduling unit 2 can directly locate the cause of the deviation based on this signal. After the trigger signal is sent, the policy update trigger module 43 starts a 1-minute timeout monitoring mechanism. If the "policy update complete" confirmation signal is not received from the AI-driven dynamic optimization scheduling unit 2 within 1 minute, the trigger signal is sent again (up to 3 times) to avoid closed-loop interruption due to communication packet loss. After receiving the trigger signal, the AI-driven dynamic optimization scheduling unit 2 will use the latest actual operating data collected by the microgrid status acquisition module 41 as input to re-iterate and calculate the scheduling strategy. After the new strategy is generated, it will be sent to the photovoltaic consumption-charging and swapping collaborative control unit 3 through the original link. The strategy update trigger module 43 will receive the identification information of the new strategy and record the update log to complete a closed-loop scheduling process.

[0067] In summary, the strategy update triggering module 43 ensures the efficiency of dynamic correction of the scheduling strategy through clear triggering logic and communication guarantee mechanism, and avoids the microgrid from becoming unstable due to the accumulation of deviations.

[0068] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0069] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids, characterized in that: include: The data acquisition and preprocessing unit (1) collects photovoltaic array output, charging and swapping load, microgrid bus electrical parameters and electricity price-related operating data in real time through distributed sensors, smart meters and microgrid monitoring system, performs noise reduction and standardization preprocessing on the collected data, generates a standardized dataset and stores it in the time series real-time database; The AI-driven dynamic optimization scheduling unit (2) is based on a standardized dataset. It constructs a photovoltaic power output prediction model through a multi-timescale long short-term memory network, outputs the photovoltaic power output prediction value and its confidence interval, and uses the confidence interval as the input state variable of the improved deep deterministic strategy gradient algorithm. It introduces a dynamic risk weight term based on the confidence interval width, adaptively adjusts the reward function and exploration rate parameters, and generates a dynamic scheduling strategy that adapts to the dynamic characteristics of the microgrid. Photovoltaic consumption-charging and swapping coordinated control unit (3), the photovoltaic consumption-charging and swapping coordinated control unit (3) receives dynamic scheduling strategy, performs orderly charging control on the charging pile, performs energy storage scheduling on the energy storage system of the swapping station, and realizes dynamic matching between photovoltaic output and charging and swapping load; The status monitoring and feedback unit (4) collects the actual operating status data of the microgrid in real time and compares the deviation with the predicted value in the dynamic scheduling strategy. When the deviation exceeds the preset deviation threshold, the AI-driven dynamic optimization scheduling unit (2) is triggered to re-iterate and update the scheduling strategy to form a closed-loop scheduling mechanism.

2. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 1, characterized in that, The data acquisition and preprocessing unit (1) includes a data acquisition module (11) and a data preprocessing and storage module (12), wherein: The data acquisition module (11) collects photovoltaic panel temperature, light intensity and photovoltaic array output power through distributed sensors, collects real-time power, current and electricity price data of charging and swapping load through smart meters, and collects microgrid bus voltage, current, total active power and reactive power through microgrid monitoring system. The data preprocessing and storage module (12) uses the 3σ criterion to remove outliers from the various raw data acquired by the data acquisition module (11), and maps the denoised data to a unified numerical range through the min-max normalization method to generate a standardized dataset. The standardized dataset is then stored in the time-series real-time database according to the dimension of "data type-acquisition time-device number" to support the subsequent AI-driven dynamic optimization scheduling unit (2) to quickly query and call the data.

3. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 2, characterized in that, The AI-driven dynamic optimization scheduling unit (2) includes a multi-timescale prediction module (21), an uncertainty quantification module (22), a risk perception reinforcement learning module (23), and a policy adaptive adjustment module (24), wherein: The multi-timescale prediction module (21) realizes short-term and medium-term prediction of photovoltaic power output based on a multi-time window long short-term memory network, and outputs the prediction confidence interval. The uncertainty quantification module (22) calculates the photovoltaic output prediction confidence interval based on the prediction feature sequence; The risk perception reinforcement learning module (23) introduces confidence interval information as input state variable into the improved deep deterministic policy gradient algorithm; The strategy adaptive adjustment module (24) dynamically adjusts the reward function weight and exploration rate parameter of the reinforcement learning algorithm according to the confidence interval in order to generate a dynamic scheduling strategy.

4. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 3, characterized in that, The process by which the multi-timescale prediction module (21) outputs the prediction confidence interval includes the following steps: S21.1 Set short-term and medium-to-long-term time windows, and construct short-term prediction sub-networks and medium-to-long-term prediction sub-networks respectively. Both short-term and medium-to-long-term prediction sub-networks adopt the long short-term memory network structure. S21.2 Input short-term photovoltaic power output historical data from the standardized dataset into the short-term prediction subnetwork, and input medium- and long-term photovoltaic power output historical data and corresponding meteorological data from the standardized dataset into the medium- and long-term prediction subnetwork. Minimize the error between the predicted value and the actual value through the gradient descent method, and iterate until the error of the validation set converges and stabilizes. S21.

3. The prediction outputs of the short-term prediction sub-network and the medium- and long-term prediction sub-network are weighted and fused through the attention weight fusion mechanism. During the fusion, the weight ratio of the outputs of the short-term prediction sub-network and the medium- and long-term prediction sub-network is dynamically allocated according to the microgrid's requirements for prediction accuracy at different time periods to obtain the preliminary photovoltaic power output prediction value. S21.4 Transmit the preliminary photovoltaic power output prediction value after fusion to the uncertainty quantification module (22).

5. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 4, characterized in that, The uncertainty quantification module (22) calculates the confidence interval for photovoltaic power output prediction using the following steps: S22.

1. The Monte Carlo dropout method is embedded in the long short-term memory network of the multi-timescale prediction module (21) to keep the dropout layer in the network training phase effective in the prediction phase. S22.

2. The input data of the multi-timescale prediction module (21) is subjected to multiple independent forward propagation through a long short-term memory network that embeds the Monte Carlo dropout method and keeps the dropout layer active, to obtain multiple parallel prediction output values. S22.3 Statistical analysis is performed on the output values ​​of multiple parallel predictions to calculate the final photovoltaic power output prediction value. and variance ; S22.4, combined with confidence coefficient Construct a confidence interval for predicting photovoltaic power output and calculate the width of the confidence interval. Confidence interval and width Output to the risk perception reinforcement learning module (23).

6. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 5, characterized in that, The process of embedding confidence interval information in the risk perception reinforcement learning module (23) includes the following steps: S23.

1. Based on the original state variables of the improved deep deterministic policy gradient algorithm, a new confidence interval width is added. As state variables, they expand the state input dimension of the algorithm; S23.2, Based on the width of the confidence interval Width of the historical maximum confidence interval The ratio is used to classify uncertainty levels and set the exploration rate. The range of values, where a first threshold is set. Second threshold ,and : when hour, Take the lower range of values; when hour, Take the middle range of values; when hour, Take the higher range of values; S23.

3. Input the expanded complete state variables into the actor network of the algorithm, so that the reinforcement learning agent can directly perceive the uncertainty of prediction during the decision-making process.

7. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 6, characterized in that, The process of the strategy adaptive adjustment module (24) optimizing algorithm parameters and generating scheduling strategies includes the following steps: S24.1, Confidence interval width based on the output of the uncertainty quantization module (22) Width of the historical maximum confidence interval Calculate dynamic risk weights ; S24.2, Dynamic risk weights Introducing a reward function to form a multi-objective technical optimization reward function ; S24.

3. Introduce a priority sampling rule and set a threshold for the empirical replay pool of the improved deep deterministic strategy gradient algorithm. ,right Scheduling experience is assigned a higher sampling weight, and the actor-evaluator network parameters of this type of experience-based update algorithm are extracted first; among which... satisfy ; S24.4 Based on the output results of the optimized improved deep deterministic strategy gradient algorithm, combined with the rated power of the microgrid photovoltaic inverter, the charge and discharge rate limit of the energy storage system, and the capacity constraint of the charging pile, a dynamic scheduling strategy including photovoltaic power distribution instructions, energy storage charge and discharge power instructions, and charging pile power scheduling instructions is generated and output to the photovoltaic consumption-charging and swapping collaborative control unit (3).

8. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 7, characterized in that, The photovoltaic consumption-charging and swapping coordinated control unit (3) includes a strategy receiving and parsing module (31), a charging terminal demand management module (32), a charging pile orderly control module (33), an energy storage system scheduling module (34), and a supply and demand matching monitoring module (35), wherein: The strategy receiving and parsing module (31) is used to receive the dynamic scheduling strategy output by the strategy adaptive adjustment module (24), and decompose the photovoltaic power distribution instruction, energy storage charging and discharging power instruction, and charging pile power scheduling instruction in the dynamic scheduling strategy into executable equipment control parameters that are adapted to the three core equipment in the microgrid: photovoltaic inverter, energy storage system and charging pile. The charging terminal demand management module (32) is used to collect the power consumption requests of the charging terminals in the charging and swapping microgrid, obtain power consumption request parameters including charging power demand and power consumption priority identifier, and convert the power consumption request parameters including charging power demand and power consumption priority identifier into load demand signals that can be recognized by the charging pile orderly control module (33). The charging pile orderly control module (33) dynamically adjusts the output power and power supply priority of each charging pile in the charging and swapping microgrid based on the maximum allowable output power parameters of the charging pile decomposed by the strategy receiving and parsing module (31) and the load demand signal of the charging terminal demand management module (32), and executes orderly charging control. The energy storage system scheduling module (34) adjusts the charging and discharging status and charging and discharging rate of the energy storage system supporting the battery swapping station in real time according to the energy storage charging and discharging current limit parameters decomposed by the strategy receiving and parsing module (31). The supply and demand matching monitoring module (35) is used to collect real-time photovoltaic actual output data, charging and swapping load real-time data and energy storage system status data, monitor the matching degree between photovoltaic output and charging and swapping load, and provide data feedback for subsequent fine-tuning of control parameters.

9. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 8, characterized in that, The process by which the photovoltaic power generation and charging / swapping coordinated control unit (3) achieves dynamic matching between photovoltaic power output and charging / swapping load includes the following steps: S31.1 Receive the dynamic scheduling strategy output by the AI-driven dynamic optimization scheduling unit (2), and decompose the photovoltaic power distribution instruction, energy storage charging and discharging power instruction, and charging pile power scheduling instruction in the strategy into executable equipment control parameters that are adapted to the photovoltaic inverter, energy storage system and each charging pile in the microgrid. S31.2 Collect the power consumption requests of charging terminals in the charging and swapping microgrid, obtain power consumption request parameters including charging power demand and power consumption priority identifier, and convert the power consumption request parameters including charging power demand and power consumption priority identifier into load demand signals that can be used for charging pile control. S31.

3. Combining the maximum allowable output power parameters of the charging pile with the load demand signal, dynamically adjust the output power and power supply priority of each charging pile, prioritize the allocation of the photovoltaic power quota to the charging terminal, and supplement the power gap through the energy storage system of the battery swapping station. S31.4 According to the charging and discharging current limit parameters of the energy storage system, when the actual output of photovoltaic power is greater than the charging and swapping load, control the energy storage system of the swapping station to perform charging operation; when the actual output of photovoltaic power is less than the charging and swapping load, control the energy storage system of the swapping station to perform discharging operation to compensate for the power gap. S31.

5. Real-time collection of actual photovoltaic output data, real-time charging and swapping load data and status data of the energy storage system supporting the swapping station, monitoring the matching degree between photovoltaic output and charging and swapping load. If the difference between the two exceeds the preset fluctuation range, the deviation data is fed back to the AI-driven dynamic optimization scheduling unit (2) for secondary optimization of the dynamic scheduling strategy.

10. The AI-driven dynamic optimization scheduling system for photovoltaic-based charging and swapping microgrids according to claim 9, characterized in that, The status monitoring and feedback unit (4) includes a microgrid status acquisition module (41), a deviation analysis module (42), and a strategy update triggering module (43), wherein: The microgrid status acquisition module (41) is used to collect the actual operating status data of the charging and swapping microgrid in real time, covering the core operating parameters of the microgrid, including the actual output of photovoltaic power, the actual consumption of charging and swapping load, the actual charging and discharging power of the energy storage system, and the actual output power of the charging pile. The deviation analysis module (42) is used to compare the actual operating status data collected by the microgrid status acquisition module (41) with the predicted value in the dynamic scheduling strategy output by the AI-driven dynamic optimization scheduling unit (2), calculate the deviation value and determine whether it exceeds the preset deviation threshold. The strategy update triggering module (43) is used to trigger the AI-driven dynamic optimization scheduling unit (2) to re-iterate and update the scheduling strategy when the deviation analysis module (42) determines that the deviation exceeds the preset deviation threshold, so as to build a closed-loop scheduling mechanism.

Images