Cooperative optimization method of microgrid in industrial park with high proportion of new energy vehicles

Abstract

This invention relates to a collaborative optimization method for microgrids in industrial parks with a high proportion of new energy vehicles, comprising the following steps: acquiring real-time monitoring data on new energy power generation, energy storage device status, and production equipment load demand within the microgrid; combining historical data to predict the new energy power generation, available charging and discharging capacity of energy storage devices, and production load demand in future time periods using a time series prediction model; based on the prediction results, constructing a mixed-integer linear programming model with the minimum overall energy cost as the optimization objective, and solving it to generate a new energy power allocation scheme, an energy storage device scheduling scheme, and a production task scheduling scheme; adjusting the allocation ratio of new energy power between production load and energy storage devices according to the new energy power allocation scheme; this method improves the efficiency of new energy utilization, optimizes energy storage device scheduling and production task execution sequence, and reduces overall energy costs, making it suitable for industrial park microgrid environments with a high proportion of new energy access.

Description

Technical Field

[0001] The present invention relates to the technical field of control and regulation of power supply, and in particular to a method for collaborative optimization of microgrids in industrial parks with a high proportion of new energy vehicles. Background Art

[0002] With the rapid development of the new energy vehicle industry, industrial park microgrids have gradually introduced large-scale new energy power generation facilities, such as solar photovoltaic and wind power generation. At the same time, in order to cope with the intermittent and volatile nature of new energy power generation, energy storage equipment is generally configured in the park to balance supply and demand. In addition, the diversity and flexibility of production equipment load requirements have also made the energy management of industrial parks gradually develop in the direction of intelligence and efficiency. Existing technologies have applied a variety of optimization algorithms to coordinate the management of energy distribution, energy storage scheduling and production tasks to improve overall energy utilization efficiency.

[0003] When dealing with a high proportion of new energy access, existing technologies usually face the following major problems: insufficient prediction accuracy of new energy generation, making it difficult to accurately capture intermittent characteristics; the scheduling optimization of energy storage equipment lacks accurate response to dynamic demand and changes in equipment characteristics, resulting in low energy storage utilization; the coordinated scheduling of production equipment loads is difficult to effectively combine with energy allocation plans, resulting in a contradiction between production task execution efficiency and energy costs. In addition, existing technologies often only optimize for a single goal, lack a systematic and global comprehensive optimization plan, and are difficult to meet the complex balance of energy and production needs in industrial parks.

[0004] In order to address the above problems, the present invention proposes a microgrid collaborative optimization method for an industrial park with a high proportion of new energy vehicles. Summary of the invention

[0005] The present application provides a microgrid collaborative optimization method for an industrial park with a high proportion of new energy vehicles, so as to reduce the energy cost of the industrial park with a high proportion of new energy vehicles.

[0006] The present application provides a method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park, including:

[0007] Obtain real-time monitoring data on renewable energy power generation, energy storage equipment status, and production equipment load demand in the microgrid, and use a time series prediction model in combination with historical data to predict renewable energy power generation, available charge and discharge capacity of energy storage equipment, and production load demand in future time periods;

[0008] Based on the predicted new energy power generation, available charge and discharge capacity of energy storage equipment and production load demand, a mixed integer linear programming model for collaborative optimization with the goal of minimizing comprehensive energy cost is constructed; the mixed integer linear programming model is solved to obtain a new energy power allocation plan, an energy storage equipment scheduling plan and a production task scheduling plan;

[0009] According to the new energy power allocation plan, the allocation ratio of new energy power between production load and energy storage equipment is adjusted; according to the energy storage equipment scheduling plan, the charging and discharging behavior of the energy storage equipment is controlled; according to the production task scheduling plan, the execution order and time window of the production tasks are adjusted, and the time and rate of the flexible load are optimized.

[0010] Furthermore, the time series prediction model includes an input layer, a feature extraction layer, an attention mechanism layer and an output layer;

[0011] The input layer is used to receive new energy power generation history data, energy storage equipment status data, production load demand data and environmental impact data, and normalize the received data to obtain a multidimensional time series tensor;

[0012] The feature extraction layer is used to receive the multidimensional time series tensor provided by the input layer, extract time series features from the received data through stacked bidirectional LSTM units, and obtain a high-dimensional time series feature vector; the stacked bidirectional LSTM units include a first layer of bidirectional LSTM, a second layer of bidirectional LSTM, and a third layer of bidirectional LSTM connected in sequence; wherein the first layer of bidirectional LSTM focuses on the characteristics of new energy power generation and captures its periodic fluctuations; the second layer of bidirectional LSTM focuses on the dynamic characteristics of energy storage equipment; the third layer of bidirectional LSTM processes production load requirements and extracts the periodic characteristics of task loads;

[0013] The attention mechanism layer is used to receive the high-dimensional time series feature vector output by the feature extraction layer, dynamically adjust the weight of the time step based on the time attention mechanism, and generate a weighted feature vector; wherein, the time attention mechanism specifically includes that for the characteristics of new energy power generation, the attention mechanism tends to highlight the impact of recent meteorological conditions; for the dynamics of energy storage equipment, the weight is mainly allocated to the historical period when the load charge changes drastically; for production load requirements, the attention mechanism emphasizes the key time steps related to the specific task period;

[0014] The output layer is used to receive the weighted feature vector output by the attention mechanism layer; the weighted feature vector is mapped using a fully connected layer to convert the high-dimensional feature vector into a prediction result, which includes a new energy power generation power prediction, an energy storage device status prediction, and a production load demand prediction; the output layer is implemented using a multi-task learning framework.

[0015] The beneficial effects of the technical solution provided by this application include:

[0016] (1) Through the time series prediction model, the future renewable energy power generation is accurately predicted. Combined with the dynamic scheduling of energy storage equipment, the new energy with large volatility is effectively absorbed, the abandonment of new energy is reduced, and the utilization rate of new energy is significantly improved. (2) By constructing a mixed integer linear programming model with the lowest comprehensive energy cost as the goal, and combining the allocation of new energy, energy storage scheduling and production task optimization, the operation of each link is coordinated to effectively reduce the overall energy operation cost of the park. (3) Through the acquisition and prediction of real-time monitoring data and the dynamic adjustment of the charging and discharging behavior of energy storage equipment, the volatility of renewable energy power generation is smoothed, and the scheduling of production tasks is optimized to avoid the impact of load peaks on the stability of the microgrid, thereby improving the stability and reliability of system operation. (4) Combined with flexible load optimization, the execution order and time window of production tasks are adjusted to make full use of the low-peak electricity price and high-proportion renewable energy power supply time period, improve the flexibility of production scheduling, and meet the balance between production efficiency and energy cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of a method for collaborative optimization of microgrids in an industrial park with a high proportion of new energy vehicles provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0018] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0019] The first embodiment of the present application provides a method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 A first embodiment of the present application provides a method for collaborative optimization of microgrids in an industrial park with a high proportion of new energy vehicles, which is described in detail.

[0020] Step S101: Acquire real-time monitoring data of renewable energy power generation, energy storage equipment status, and production equipment load demand in the microgrid, and use a time series prediction model in combination with historical data to predict renewable energy power generation, available charge and discharge capacity of energy storage equipment, and production load demand in future time periods.

[0021] Step S101 is a basic step of the collaborative optimization method for microgrids in high-proportion new energy vehicle industrial parks, and is used to achieve real-time acquisition and prediction of core data in the microgrid, thereby providing key input for subsequent steps. The following is a detailed implementation description of step S101.

[0022] First, the data of renewable energy power generation is collected in real time through the monitoring system of the microgrid. This data usually comes from the output power monitoring module of the solar photovoltaic power generation equipment and wind power generation equipment in the microgrid, recording its power changes within a certain time interval (such as every 15 minutes or every hour). The integrity and time sequence of the data need to be ensured during collection to accurately reflect the fluctuation characteristics of renewable energy power generation.

[0023] Secondly, obtain the status data of the energy storage device, including the current charge and discharge power, charge and discharge efficiency and status parameters of the energy storage device (such as load state SOC, State of Charge, i.e. the current power state of the energy storage device). This data is provided by the energy storage management system (BMS) or energy storage controller and recorded in the form of time series to ensure that it dynamically reflects the operation and historical status of the energy storage device.

[0024] Subsequently, the load demand data of production equipment is obtained from the production management system in the industrial park. This data usually includes the load demand curve of the production task, the task execution time window, and the operating power consumption of the equipment. The production management system may also provide classified information about flexible loads (i.e., tasks whose load demand can be adjusted) and rigid loads (i.e., tasks that must be completed on time) for subsequent optimization.

[0025] On the basis of real-time monitoring data acquisition, historical data is combined with real-time data, and a time series prediction model is used to predict the renewable energy power generation, available charge and discharge capacity of energy storage equipment, and production load demand in the future time period. The time series prediction model can be implemented based on a recursive neural network (such as LSTM, long short-term memory network) or a statistical method (such as ARIMA model, autoregressive integral moving average model). The model input includes historical power data of renewable energy power generation, SOC change trend of energy storage equipment, historical load demand of production tasks, and environmental influencing factors related to these data (such as meteorological data, fluctuations in purchased electricity prices).

[0026] When making predictions, historical data is divided into training sets and validation sets, and model parameters are adjusted based on training data to ensure that the model has high accuracy in predicting future periods. For the prediction of renewable energy power generation, the changing trend of meteorological conditions (such as light intensity and wind speed) is taken into consideration; for the prediction of available charge and discharge capacity of energy storage equipment, the historical fluctuation of SOC and rated performance of equipment are combined; for the prediction of production load demand, the possible fluctuation of load demand in the future is analyzed based on task planning and historical equipment operation data.

[0027] After the prediction is completed, three time series prediction results are generated: the new energy power generation curve in the future time period, the charging and discharging capacity change trend of the energy storage equipment, and the load demand curve of the production task. These prediction results will serve as the core input data for the subsequent step S102 to build the optimization model, supporting the collaborative optimization of new energy power allocation, energy storage equipment scheduling and production task scheduling. Through the above steps, it is ensured that the implementation of step S101 meets the requirements of data real-time and prediction accuracy, laying the foundation for the realization of the entire collaborative optimization method.

[0028] Furthermore, the time series prediction model includes an input layer, a feature extraction layer, an attention mechanism layer and an output layer;

[0029] The input layer is used to receive new energy power generation history data, energy storage equipment status data, production load demand data and environmental impact data, and normalize the received data to obtain a multidimensional time series tensor;

[0030] The feature extraction layer is used to receive the multidimensional time series tensor provided by the input layer, extract time series features from the received data through stacked bidirectional LSTM units, and obtain a high-dimensional time series feature vector; the stacked bidirectional LSTM units include a first layer of bidirectional LSTM, a second layer of bidirectional LSTM, and a third layer of bidirectional LSTM connected in sequence; wherein the first layer of bidirectional LSTM focuses on the characteristics of new energy power generation and captures its periodic fluctuations; the second layer of bidirectional LSTM focuses on the dynamic characteristics of energy storage equipment; the third layer of bidirectional LSTM processes production load requirements and extracts the periodic characteristics of task loads;

[0031] The attention mechanism layer is used to receive the high-dimensional time series feature vector output by the feature extraction layer, dynamically adjust the weight of the time step based on the time attention mechanism, and generate a weighted feature vector; wherein, the time attention mechanism specifically includes that for the characteristics of new energy power generation, the attention mechanism tends to highlight the impact of recent meteorological conditions; for the dynamics of energy storage equipment, the weight is mainly allocated to the historical period when the load charge changes drastically; for production load requirements, the attention mechanism emphasizes the key time steps related to the specific task period;

[0032] The output layer is used to receive the weighted feature vector output by the attention mechanism layer; the weighted feature vector is mapped using a fully connected layer to convert the high-dimensional feature vector into a prediction result, which includes a new energy power generation power prediction, an energy storage device status prediction, and a production load demand prediction; the output layer is implemented using a multi-task learning framework.

[0033] Furthermore, the input layer is specifically used for:

[0034] Receive new energy power generation history data, energy storage equipment status data, production load demand data and environmental impact data, and format the received data into a unified timestamp sequence;

[0035] The formatted data is normalized, where the historical data of renewable energy power generation is based on the maximum and minimum power in the past year, the status data of energy storage equipment is based on the rated capacity and maximum power of the equipment, the production load demand data is based on the maximum load demand of the equipment, and the environmental impact data is normalized based on the historical meteorological data range;

[0036] The normalized time series is divided into time segments of fixed length through the sliding window mechanism, and a multidimensional time series tensor is generated as the input of the feature extraction layer.

[0037] The input layer first receives data from multiple sources through the microgrid data acquisition module, including historical data of renewable energy power generation, energy storage equipment status data, production load demand data, and environmental impact data. The historical data of renewable energy power generation usually comes from the monitoring equipment of solar photovoltaic and wind power generation, which records the historical power change curve, and the time interval can be set to minutes or hours; the energy storage equipment status data is provided by the energy storage management system (BMS), including historical charge and discharge power, equipment SOC (State of Charge), charge and discharge efficiency and other dynamic status information; the production load demand data comes from the production scheduling system, covering the historical load curve of production tasks, task execution time and flexible load characteristics; the environmental impact data usually comes from the meteorological monitoring system, including the fluctuation history of light intensity, wind speed, temperature and purchased electricity price.

[0038] The received data may be recorded in different time formats and time intervals. The input layer needs to format these data uniformly into a time series with consistent timestamps. The formatting process includes time alignment and completion operations on the data. For example, for data with inconsistent time intervals, interpolate or aggregate them into a uniform time step (such as 15 minutes or 1 hour) to ensure data integrity and consistency of the time series. Accurate alignment of timestamps is crucial for subsequent normalization and modeling steps.

[0039] After data formatting, the input layer normalizes the data to eliminate scale differences between different data features, thereby improving the training efficiency and prediction accuracy of the model. The normalization of historical data of renewable energy generation is based on the maximum power and minimum power in the past year, mapping all data points to the range of [0,1]; the energy storage device status data is normalized based on the rated capacity and maximum power of the device to ensure that state information such as SOC meets the input requirements of the model; production load demand data is normalized according to the maximum load demand of the equipment to ensure that the relative weight of the task load is retained; environmental impact data (such as meteorological conditions) is based on the range of historical meteorological data and adopts the same normalization strategy.

[0040] The normalized time series data is further divided into fixed-length time segments through the sliding window mechanism. The function of the sliding window is to divide the continuous time series into data blocks with overlapping or non-overlapping windows. For example, the window length is set to 24 time steps (corresponding to one day's data) and the sliding step is 1 time step. Each time segment contains a fixed number of time steps to ensure that the short-term dynamic characteristics and long-term trends of the time series can be captured.

[0041] The time segments generated by the sliding window are organized into a multidimensional time series tensor. The structure of the tensor usually contains three dimensions: time step, feature dimension, and number of samples. Each time step corresponds to a specific time point of a time segment. The feature dimension includes multiple features such as renewable energy power generation, energy storage equipment status, production load demand, and environmental conditions. The number of samples represents the total number of time segments generated by the sliding window. The multidimensional time series tensor generated in the end is used as the output of the input layer, providing standardized, structured, and high-quality data input for the subsequent feature extraction layer.

[0042] Furthermore, the first-layer bidirectional LSTM is specifically used for:

[0043] Receive the normalized renewable energy power generation data and meteorological condition data (including light intensity, wind speed, temperature, etc.), and process the data through a bidirectional LSTM unit to capture the periodic fluctuation characteristics of renewable energy power generation; the forward LSTM unit is used to capture the increasing trend of renewable energy power generation characteristics over time, and the reverse LSTM unit captures the degradation pattern of historical data; combine the outputs in the forward and reverse directions to generate a complete renewable energy power generation time series feature vector as the input of the second layer of bidirectional LSTM.

[0044] First, the normalized renewable energy power generation data and related meteorological condition data are received from the input layer. These data are usually expressed in the form of time series, covering multiple continuous time steps. The renewable energy power generation data contains historical solar photovoltaic and wind power generation values. After normalization, their values ​​are standardized to a specific range, such as [0,1], to eliminate the impact of data magnitude. Meteorological condition data include light intensity, wind speed, and temperature, which directly affect the output power of renewable energy power generation. In order to ensure the consistency of the time series, the meteorological condition data is also formatted and normalized, corresponding to the renewable energy power generation data one by one.

[0045] The bidirectional LSTM unit is used to process these time series data, focusing on capturing the periodic fluctuation characteristics of the data. LSTM (Long Short-Term Memory Network) is a type of recurrent neural network that can effectively model long-term dependencies in time series. The bidirectional LSTM consists of two LSTM units in opposite directions, which process the forward and reverse information of the time series respectively. The forward LSTM unit reads data in sequence along the forward direction of the time series to extract the increasing trend of renewable energy power generation and meteorological conditions over time, such as the power generation increase pattern brought about by the gradual increase in daytime light. The reverse LSTM unit reads data from the tail of the time series in reverse to capture the degradation pattern in the historical data, such as the power generation decrease pattern caused by the gradual decrease in light intensity.

[0046] Through this bidirectional processing method, the forward and reverse LSTM units extract the complete characteristics of the time series, making up for the reverse information that may be ignored by the unidirectional LSTM, thereby more comprehensively capturing the periodic fluctuations of renewable energy power generation. The outputs of the two LSTM units are combined at each time step to generate a complete time series feature vector. This feature vector not only contains the changing trend of renewable energy power data over time, but also combines the dynamic impact of meteorological conditions on power generation capacity, which can reflect the periodic laws of photovoltaic and wind power generation.

[0047] The generated time series feature vectors of renewable energy generation are organized into structured high-dimensional data as the input of the second-layer bidirectional LSTM, providing a basis for the subsequent dynamic characteristic modeling of energy storage equipment. This layer-by-layer progressive feature extraction method ensures the independence and hierarchical correlation of different data characteristics in the model, providing a high-quality input basis for the entire microgrid collaborative optimization method.

[0048] Furthermore, the second-layer bidirectional LSTM is specifically used for:

[0049] The new energy power generation feature vector from the first layer of bidirectional LSTM is received, and combined with the historical SOC data, charging and discharging power data, and charging and discharging efficiency data of the energy storage device, the input data is processed through the bidirectional LSTM unit to capture the charging and discharging behavior rules and SOC change patterns in the dynamic characteristics of the energy storage device; among them, the forward LSTM unit focuses on the trend prediction of the future charging and discharging behavior of the energy storage device, and the reverse LSTM unit captures the historical charging and discharging abnormal changes, and generates the time series feature vector of the dynamic characteristics of the energy storage device through the fusion of features in the positive and negative directions as the input of the third layer of bidirectional LSTM.

[0050] The second-layer bidirectional LSTM first receives the new energy generation feature vectors generated by the first layer. These feature vectors already contain information about the periodic fluctuations in power generation and the dynamic impact of meteorological conditions on power generation. At the same time, relevant data of the energy storage device is obtained from the energy storage management system, including SOC (state of charge and discharge) data, historical charge and discharge power, and charge and discharge efficiency. SOC data reflects the current and historical charge level of the energy storage device, usually expressed as a percentage, ranging from 0% to 100%. The charge and discharge power data describes the power output or input of the energy storage device at different time steps, in kilowatts (kW), which can reflect the operating load of the energy storage device. The charge and discharge efficiency data represents the proportion of loss of the energy storage device during the energy conversion process, which is used to characterize the operating efficiency of the device.

[0051] After the above data is combined, it is input into the bidirectional LSTM unit for processing. The forward LSTM unit processes data from the past to the future in the order of time series, focusing on analyzing the charging and discharging behavior of energy storage equipment. For example, when the power of renewable energy generation is sufficient, the SOC of energy storage equipment usually shows an upward trend, while when the power generation is insufficient, the decrease in SOC may be directly related to the production load demand. The forward LSTM also predicts the future operation mode of energy storage equipment through historical trends, such as possible charging peaks or discharging troughs.

[0052] The reverse LSTM unit processes data from the future to the past in the reverse order of the time series, focusing on capturing abnormal changes in historical data. For example, an abnormal and significant drop in SOC at certain time steps may be related to high-load discharge or equipment operation failure. Through reverse analysis, the reverse LSTM can help the model better understand the potential impact of these historical anomalies on the current and future energy storage status.

[0053] The features extracted by the LSTM units in the forward and reverse directions are fused in the time dimension to form a comprehensive time series feature vector of the dynamic characteristics of the energy storage device. These feature vectors not only contain the operation trend prediction of the energy storage device in the future time period, but also combine the historical charging and discharging behavior and the abnormal characteristics of SOC changes, thus providing comprehensive input for the subsequent optimization steps.

[0054] The final generated time series feature vector of the dynamic characteristics of the energy storage device is passed to the third layer of bidirectional LSTM to further model the characteristics related to production load demand. This design ensures the effective integration of the characteristics of new energy generation and the dynamic characteristics of energy storage equipment, and improves the pertinence and accuracy of feature extraction through hierarchical modeling.

[0055] Furthermore, the third-layer bidirectional LSTM is specifically used for:

[0056] The dynamic feature vector of the energy storage device is received from the second-layer bidirectional LSTM, and combined with the historical curve of the production load demand in the industrial park, the task scheduling plan, and the task completion time, the input data is processed through the bidirectional LSTM unit to capture the periodic variation characteristics of the production load demand; among them, the forward LSTM unit analyzes the gradual growth pattern of the load demand with the start and end of the task, and the reverse LSTM unit focuses on the sudden load and peak load between historical tasks, and merges the results in the forward and reverse directions into a high-dimensional time series feature vector of the comprehensive production load demand as the input of the attention mechanism layer.

[0057] The third-layer bidirectional LSTM first receives the dynamic feature vectors of the energy storage device output from the second-layer bidirectional LSTM. These feature vectors already contain the dynamic characteristics of the energy storage device in terms of charging and discharging behavior and SOC changes, and also integrate the characteristics of new energy power generation. In addition, this layer also receives a variety of data inputs related to production load demand in the industrial park. The historical curve of production load demand provides basic information about the change of load demand over time, usually expressed in the form of a time series, covering past task execution records. The task scheduling plan includes the start time and expected completion time of each production task, clarifying the time window and resource requirements of the production task. The historical data of task completion time reflects whether the task is in line with the plan during actual execution, providing an important basis for predicting future task loads.

[0058] After these input data are combined, they are input into the bidirectional LSTM unit for processing. The forward LSTM unit processes the data in the forward order of the time series, focusing on analyzing the gradual load growth pattern of the production task from start to end. For example, when the task is close to the start time, the load demand may gradually increase, and the load demand may gradually decrease at the end of the task. Through this trend analysis, the forward LSTM can identify the typical change pattern of load demand during the execution of the production task and predict the load trend of future tasks.

[0059] The reverse LSTM unit processes data in reverse order of the time series, focusing on the load variation characteristics between historical tasks, such as the time when sudden loads occur and the specific conditions of peak loads. These characteristics usually appear during the execution of flexible tasks or when task plans are changed due to emergencies. By reverse modeling these historical data, the reverse LSTM can help identify the impact of past load anomalies on current and future production load requirements.

[0060] The features extracted by the LSTM units in the forward and reverse directions are fused in the time dimension to generate a comprehensive high-dimensional production load demand time series feature vector. This feature vector contains both the time-period law of task execution and the dynamic characteristics of burst load and peak load, so as to fully describe the production load demand of the industrial park.

[0061] The final generated comprehensive production load demand feature vector is passed to the attention mechanism layer to further strengthen the dynamic weight allocation of key time steps. By modeling the dynamic characteristics of energy storage equipment and production load demand characteristics layer by layer, the third layer of bidirectional LSTM ensures the complete expression of data characteristics and provides high-quality input for the collaborative optimization of microgrids.

[0062] Furthermore, the output layer is specifically used for:

[0063] Receive the weighted feature vectors from the attention mechanism layer, and assign the weighted feature vectors to different task prediction branches according to the characteristics of renewable energy power generation, the dynamic characteristics of energy storage equipment, and the characteristics of production load demand;

[0064] In the renewable energy power generation prediction branch, weighted features related to meteorological conditions are used to further process environmental data including light intensity and wind speed, and the renewable energy power generation prediction results for the future time period are output;

[0065] In the energy storage device state prediction branch, based on the historical load state, charge and discharge efficiency and dynamic characteristic weighted features of the energy storage device, combined with the impact of future task scheduling load on the charge and discharge behavior of the energy storage device, the available charge and discharge capacity and state of the energy storage device in the future time period are output;

[0066] In the production load demand forecasting branch, the weighted features related to the production task period are used, combined with the historical task start and end time and load trends, to generate a production task load demand forecast for the future time period;

[0067] Through the multi-task learning framework, the underlying features shared by the above prediction branches are separated from the specific features of each branch, and the joint loss function is used to optimize the prediction accuracy of each task, ensuring that the prediction results of renewable energy power generation, energy storage equipment status and production load demand are coordinated with each other to meet the needs of microgrid collaborative optimization.

[0068] The output layer first receives the weighted feature vectors from the attention mechanism layer. These feature vectors have strengthened the importance analysis of different time steps through the temporal attention mechanism and contain key information about the characteristics of renewable energy generation, the dynamic characteristics of energy storage equipment, and the characteristics of production load demand. The output layer analyzes these weighted features and assigns them to different task prediction branches so that each prediction target can be processed specifically.

[0069] In the renewable energy power generation prediction branch, the output layer uses the meteorological-related parts of the weighted features to further process the dynamic changes of environmental data such as light intensity and wind speed. These environmental data directly affect the output power of renewable energy power generation. Through in-depth analysis of these features, the model can accurately predict the renewable energy power generation in the future time period. The prediction results are in the form of a time series, including the power generation value of each time step, which is used to guide the renewable energy power allocation of the microgrid.

[0070] In the energy storage device state prediction branch, the output layer combines the historical load state, charge and discharge efficiency, and dynamic characteristics of the energy storage device with weighted features, and considers the potential impact of future task scheduling loads on the energy storage device to predict the available charge and discharge capacity and operating status of the energy storage device in the future time period. This prediction branch pays special attention to the SOC (charge and discharge state) and charge and discharge power change trends of the energy storage device to ensure that the prediction results can reflect the dynamic characteristics of the energy storage device in future operation and provide a basis for energy storage scheduling optimization.

[0071] In the production load demand forecasting branch, the output layer uses weighted features related to the production task period, combined with the start and end time of historical tasks and the change trend of load demand, to generate a production task load demand forecast for the future time period. The model pays special attention to the rigid load and flexible load characteristics of production tasks. The forecast results include the specific load demand values ​​of the task in each time period, which are used to guide the scheduling optimization of production tasks.

[0072] Through the multi-task learning framework, the output layer optimizes the above three prediction branches while sharing the underlying features to improve the overall modeling efficiency. In this framework, the underlying features include the common dynamic characteristics of renewable energy generation, energy storage equipment and production loads, while the specific features of each branch focus on the corresponding prediction task. The joint loss function is used to optimize the output accuracy of the three prediction branches at the same time, and by dynamically adjusting the loss weights of each task, it ensures that their prediction results are coordinated with each other during the collaborative optimization process without conflict.

[0073] Finally, the prediction results generated by the output layer include the power generation of renewable energy in the future time period, the available charging and discharging capacity and status of energy storage equipment, and the load demand of production tasks. These results serve as input data and directly serve the coordinated optimization of microgrids to ensure energy utilization efficiency and system operation stability under the condition of high proportion of renewable energy access.

[0074] Furthermore, the attention mechanism layer is specifically used for:

[0075] Receive the high-dimensional time series feature vector output by the feature extraction layer, and divide it into feature sub-vectors corresponding to the new energy power generation characteristics, the dynamic characteristics of the energy storage equipment, and the production load demand characteristics;

[0076] For the characteristics of renewable energy power generation, the correlation between each time step and the current time is calculated based on the historical change trend of meteorological conditions, and the time steps with significant changes in light intensity and wind speed conditions are highlighted, and their impact on renewable energy power generation prediction is emphasized by assigning higher weights;

[0077] For the dynamic characteristics of energy storage equipment, analyze the historical load state changes of energy storage equipment and the periods of drastic changes in charging and discharging power, identify the time steps that are closely related to the future energy storage state, assign higher weights to the identified time steps, and enhance the accuracy of the model's prediction of the energy storage equipment state;

[0078] For production load demand characteristics, based on the scheduling plan of production tasks, key time steps close to the task start time or end time are identified and assigned higher weights, so that the characteristics of these time steps play a more important role in load demand forecasting;

[0079] The weight assigned to each time step is weighted with the corresponding characteristic sub-vector to generate weighted characteristic vectors of various characteristics; the weighted characteristic vectors of new energy power generation, energy storage dynamics and production load demand are fused into weighted characteristic vectors.

[0080] The attention mechanism layer first receives the high-dimensional time series feature vectors output by the feature extraction layer. These feature vectors are the output data processed by the aforementioned bidirectional LSTM network, and contain the multi-dimensional time series characteristics of the microgrid operation. After receiving, the attention mechanism layer divides the high-dimensional time series feature vector into three sub-vectors, corresponding to the characteristics of new energy power generation, the dynamic characteristics of energy storage equipment, and the production load demand characteristics. This division is based on the dimension and physical meaning of the input features. For example, the characteristics of new energy power generation are mainly composed of the power change characteristics of photovoltaic and wind power generation, the dynamic characteristics of energy storage equipment are mainly composed of charging and discharging states and power characteristics, and the production load demand characteristics contain the load demand time series of task execution.

[0081] For the characteristics of renewable energy power generation, the attention mechanism layer calculates the relevance weight of each time step by analyzing the historical trend of meteorological conditions. Specifically, changes in meteorological conditions, such as fluctuations in light intensity and wind speed, have a periodic characteristic of affecting renewable energy power generation. The attention mechanism assigns higher weights based on these time steps with significant changes, such as time points when light intensity gradually increases or wind speed increases significantly, so that these time steps make a greater contribution to renewable energy power generation forecasts. In this way, the model can focus on key time steps more accurately and improve the ability to predict future power generation.

[0082] For the dynamic characteristics of energy storage devices, the attention mechanism layer focuses on analyzing the historical load state changes of energy storage devices, especially the fluctuations of SOC (charge and discharge state) and the periods of drastic changes in charge and discharge power. These drastic changes usually reflect the important behaviors of energy storage devices in past operations, such as moments of rapid charging or high-load discharge. Through the calculation of the attention mechanism, these key time steps that have a greater impact on the future energy storage state are identified and assigned higher weights. In this way, the model can more effectively capture the operating laws of energy storage devices and enhance the accuracy of future energy storage state predictions.

[0083] For the production load demand characteristics, the attention mechanism layer combines the scheduling plan of the production task and focuses on identifying time steps close to the task start time or end time. These key time steps usually play an important role in the production load demand forecast. For example, the load demand may gradually increase before the task starts, and gradually decrease after the task ends. By assigning higher weights to these key time steps, the attention mechanism can highlight the periodic change characteristics of the load demand during the task scheduling process, providing support for more accurate production task load demand forecasting.

[0084] After the weight calculation of each time step is completed, the attention mechanism layer weights these weights with the corresponding feature sub-vectors to generate a weighted feature vector. The weighted processing highlights the importance of key time steps in the feature vector by dynamically adjusting the contribution of each time step. The weighted feature vectors represent the enhanced characteristics of renewable energy power generation, energy storage equipment dynamics, and production load demand. Finally, these weighted feature vectors are fused into a comprehensive weighted feature vector, which retains the important information of the three characteristics in the time dimension and forms a unified output for further processing in subsequent layers.

[0085] The following is the reference code for the time series forecasting model:

[0086]

[0087]

[0088]

[0089] Training a time series forecasting model starts with data preparation, including data loading, preprocessing, and partitioning into training, validation, and test sets. Subsequently, the model's loss function and optimizer need to be defined, and then the model's parameters are gradually optimized in an iterative process until convergence. The following are the complete training steps.

[0090] First, prepare the data. Load the renewable energy power generation data, energy storage equipment status data, and production load demand data, which are usually stored in CSV files or databases. The raw data can be converted into a matrix form through Python's Pandas or NumPy libraries. Subsequently, the data is divided in chronological order to generate fixed-length time segments, which will be used as input to the model. To ensure that the data is suitable for the deep learning model, the input data needs to be normalized, such as mapping the values ​​to the range of [0,1] to eliminate the magnitude differences between the features. The normalization criteria can be based on the historical maximum and minimum values ​​of the data. The processed data is divided into a training set, a validation set, and a test set. The training set is used for model training, the validation set is used for hyperparameter adjustment, and the test set is used to evaluate model performance.

[0091] Next, define the model structure. In the code, the time series prediction model has been implemented, which includes an input layer, a three-layer bidirectional LSTM, an attention mechanism layer, and a multi-task output layer. The model needs to be initialized and hyperparameters, such as input dimension, hidden layer dimension, sequence length, etc., need to be set. Choose an appropriate loss function to measure the error between the predicted value and the true value. For the multi-task learning framework, a separate loss function can be defined for each task, such as the mean square error (MSE), and then the total loss is obtained by weighted summation. The choice of weights can be adjusted according to the importance of the task to balance the contribution of each task.

[0092] The training process is implemented through PyTorch's optimizer. Choose Adam or SGD as the optimizer, these methods usually have good convergence effect. Before training, transfer the model and data to a device that supports accelerated computing, such as a GPU. In each training cycle, a batch of data is randomly extracted from the training set, input into the model for forward propagation, and the output of the model is calculated. Then, the total loss is calculated based on the true value and the predicted value, the gradient is calculated through the backpropagation algorithm, and finally the optimizer is used to update the model parameters. In order to avoid overfitting, Dropout layers or other regularization methods can be used during training, or training can be stopped early based on the performance of the validation set.

[0093] After each training cycle, the validation set is used to evaluate the performance of the model, and indicators such as loss value and prediction accuracy are recorded. During the training process, the validation set loss is monitored to see if it continues to decrease. If the validation loss no longer decreases within a few rounds, the training can be terminated. In order to speed up the training and prevent numerical instability, a learning rate scheduler can be set to gradually reduce the learning rate as the training progresses, so as to accurately adjust the model parameters in the later stages of optimization.

[0094] After training is complete, save the model's parameters and structure for evaluation and deployment in a test set or actual production environment. You can use the torch.save method provided by PyTorch to save the model, and use torch.load to load the model for inference. In order to evaluate the final performance of the model, test the prediction results of the three tasks in the test set, calculate various indicators such as mean square error, mean absolute error or R square value, and comprehensively analyze the advantages and disadvantages of the model.

[0095] Through the above steps, the entire time series prediction model development process can be implemented from raw data to the trained model.

[0096] Step S102: Based on the predicted new energy power generation, available charge and discharge capacity of energy storage equipment, and production load demand, a mixed integer linear programming model for collaborative optimization with the goal of minimizing the comprehensive energy cost is constructed; the mixed integer linear programming model is solved to obtain a new energy power allocation plan, an energy storage equipment scheduling plan, and a production task scheduling plan.

[0097] The main purpose of step S102 is to construct a mixed integer linear programming model (MILP) based on the time series predicted new energy power generation, the available charge and discharge capacity of the energy storage equipment, and the production load demand, and generate an optimization solution by solving the model. Step S102 is described in detail below.

[0098] First, the predicted data needs to be structured and the time series data of renewable energy power generation, available charging and discharging capacity of energy storage equipment, and production load demand needs are converted into input parameters of the optimization model. These data represent the energy supply in the microgrid, the dispatching capacity of energy storage equipment, and the energy demand of production equipment in the future time period. In order to adapt the mixed integer linear programming model, the time series data needs to be divided into discrete time steps, and the predicted value of each time step is used as the input of the corresponding variable in the model.

[0099] Next, based on the goal of minimizing the comprehensive energy cost, define the objective function of the optimization model. This objective function usually consists of three main cost items, including the cost of purchased electricity, the loss of unused new energy, and the operating cost of energy storage equipment. For example, the cost of purchased electricity can be calculated based on the price curve of the external electricity market and the actual demand for purchased electricity; the loss of unused new energy can be quantified by the difference between the predicted new energy power generation and the actual power used; the operating cost of energy storage equipment usually consists of the efficiency loss during the charging and discharging process and the loss cost related to the equipment life.

[0100] On the basis of the definition of the objective function, the constraints need to be clarified to ensure that the model solution meets the physical characteristics and operation requirements of the microgrid. The use of renewable energy power generation must be limited by its predicted supply, the charging and discharging power and state of charge (SOC) of the energy storage device must be within the rated range of the device, and the charging and discharging process cannot occur simultaneously at the same time step. The production load demand needs to match the actual production task scheduling. The rigid load must be completed within the specified time window, while the flexible load can adjust the execution time and rate during the optimization process. In addition, in order to ensure the dynamic balance between energy supply and demand, the model also needs to introduce the power balance constraint of the microgrid, that is, in each time step, the total energy supply (including renewable energy generation, energy storage discharge and purchased electricity) must be equal to the total energy demand (including production load and energy storage charging).

[0101] After the model is built, a professional solver is used to solve the mixed integer linear programming model. Based on the objective function and constraints, the solver will find the optimal solution through iterative optimization and generate optimization results. The optimization results usually include the new energy power allocation plan, energy storage equipment scheduling plan and production task scheduling plan for each future time step. Among them, the new energy power allocation plan clarifies the specific allocation ratio of new energy power generation between production load and energy storage equipment; the energy storage equipment scheduling plan defines the charging and discharging state and power of the energy storage equipment at each time step; the production task scheduling plan provides the operation time arrangement of production equipment and load demand adjustment suggestions.

[0102] Finally, the optimization solutions obtained will be output in a structured manner to provide clear guidance for the implementation of subsequent steps. These solutions will be directly used to adjust the energy distribution, energy storage scheduling and production task execution sequence of the microgrid to ensure that the microgrid can maximize energy utilization efficiency and minimize comprehensive costs under the condition of high proportion of new energy access.

[0103] Furthermore, the objective function of the mixed integer linear programming model is shown in the following formula (1):

[0104]

[0105] Among them, C total represents the total cost; C grid represents the cost of purchased electricity; C storage Represents the operating cost of energy storage equipment; C task represents the production task scheduling cost; C renewable represents the penalty cost for not using new energy; the cost of purchasing electricity C grid It is expressed using the following formula (2): λ1, λ2, λ3 and λ4 are weight coefficients; ∈ is a small positive number used to avoid the denominator being 0;

[0106] The values ​​of weight coefficients λ1, λ2, λ3, and λ4 can be dynamically adjusted according to the priority objectives of system operation. For example, in a scenario where new energy accounts for a high proportion, the weight of λ4 can be increased to strengthen the incentive for the use of new energy.

[0107]

[0108] Where T represents the total number of discrete time steps in the optimization period; P grid,t represents the purchased power at time t; λ grid represents the unit cost of purchased electricity;

[0109] The energy storage equipment operating cost C storage Use the following formula (3) to express:

[0110]

[0111] Where T represents the total number of discrete time steps in the optimization period; P charge,t represents the charging power of the energy storage device at time t; P discharge,t represents the discharge power of the energy storage device at time t; η charge Represents the charging efficiency of the energy storage device; η discharge Represents the discharge efficiency of the energy storage device; storage Indicates the unit energy storage equipment operating cost;

[0112] Among them, the production task scheduling cost C task It is expressed using the following formula (4):

[0113]

[0114] Where T represents the total number of discrete time steps in the optimization period; N represents the total number of production tasks; L i,t represents the load demand of production task i at time t; α i represents the priority weight of task i, which is determined by the production scheduling strategy; represents the actual scheduling time of task i at time t; represents the optimal scheduling time of task i at time t; λ task It represents the unit dispatch cost weight, which is determined by experimental data or directly set based on expert knowledge;

[0115] The penalty cost C for not utilizing the new energy renewable Use the following formula (5) to express:

[0116]

[0117] Where T represents the total number of discrete time steps in the optimization period; The maximum power of renewable energy generation at time t; Indicates the actual power used by renewable energy generation at time t; renewable It indicates the penalty cost weight for the unit not utilizing new energy.

[0118] Furthermore, the constraints of the mixed integer linear programming model include:

[0119] (a) Power balance constraint, expressed as follows:

[0120]

[0121] Among them, P gen,t Represents the renewable energy power generation at time t; P grid,t represents the purchased power at time t; P discharge,t Represents the energy storage discharge power at time t; P load,t represents the production load demand at time t; P charge,t represents the energy storage charging power at time t; T represents the total number of discrete time steps in the optimization time period;

[0122] (b) Energy storage device state constraint, expressed using the following formula (7):

[0123]

[0124] Among them, SOC t+1 Indicates the state of charge of the energy storage device at time t+1; SOC t represents the state of charge of the energy storage device at time t; Δt is the time step; η charge Represents the charging efficiency of the energy storage device; P charge,t represents the charging power of the energy storage device at time t; P discharge,t represents the discharge power of the energy storage device at time t; η discharge represents the discharge efficiency of the energy storage device; T represents the total number of discrete time steps in the optimization period;

[0125] (c) The operating range constraints of the energy storage device are expressed using the following formulas (8) and (9):

[0126]

[0127] Among them, SOC t SOC is the state of charge of the energy storage device at time t, indicating the current power level of the energy storage device; min SOC is the minimum allowable state of charge of the energy storage device, indicating the minimum power limit of the energy storage device when discharging; min P is the highest state of charge of the energy storage device, indicating the maximum power limit of the energy storage device when charging; charge,t represents the charging power of the energy storage device at time t; Indicates the maximum charging power of the energy storage device; P discharge,t represents the discharge power of the energy storage device at time t; Indicates the maximum discharge power of the energy storage device;

[0128] (d) The constraint condition of new energy power fluctuation is expressed by the following formula (10):

[0129]

[0130] Among them, P gen,t is the renewable energy power generation at time t; is the maximum power generation of renewable energy at time t; P gen,t+1 is the renewable energy power generation at time t+1; R ramp is the power ramp limit;

[0131] (e) The production task scheduling constraints are expressed using the following formulas (11) and (12):

[0132]

[0133] Among them, T i,start Indicates the start time of task i; Indicates the earliest start time of task i; represents the latest start time of task i; L i,t represents the load requirement of task i at time i; L i,max represents the maximum load requirement of task i; T i,start represents the actual start time of task i; T i,end represents the actual end time of task i; T represents the total number of discrete time steps in the optimization period; N represents the total number of tasks;

[0134] (f) The external power purchase restriction constraint is expressed using the following formula (13):

[0135]

[0136] Among them, P grid,t is the power purchased from the external grid at time t; is the maximum power supply of the external power grid; T represents the total number of discrete time steps in the optimization period;

[0137] (g) The renewable energy priority consumption constraint is expressed using the following formulas (14) and (15):

[0138]

[0139] Among them, P load,t represents the production load demand at time t; P charge,t represents the charging power of the energy storage device at time t; P gen,t is the renewable energy power generation at time t; P grid,t is the power purchased from the external grid at time t; It represents the actual power used by renewable energy generation at time t; T represents the total number of discrete time steps in the optimization period.

[0140] Step S103: According to the new energy power allocation plan, adjust the allocation ratio of new energy power between production load and energy storage equipment; according to the energy storage equipment scheduling plan, control the charging and discharging behavior of the energy storage equipment; according to the production task scheduling plan, adjust the execution order and time window of the production task, and optimize the time and rate of the flexible load.

[0141] Step S103 is to perform key operations of new energy power allocation, energy storage equipment scheduling and production task scheduling according to the optimization plan generated in the previous step, aiming to achieve microgrid coordinated optimization under the condition of high proportion of new energy access. Step S103 is described in detail below.

[0142] First, according to the new energy power allocation scheme, the new energy power generation in the microgrid is allocated between the production load and the energy storage device. In each time step, it is necessary to compare the predicted new energy power generation with the total load demand of the current production equipment in real time. If the new energy power generation is greater than the production load demand, the excess power generation will be used first to charge the energy storage device until the SOC (State of Charge, load state, also called charge and discharge state) of the energy storage device reaches the upper limit of its rated capacity. If the new energy power generation is not enough to meet the production load demand, it is necessary to supplement it in combination with the discharge power of the energy storage device, and purchase electricity from the external power grid when necessary to ensure the stable energy supply of the production load. In this process, it is necessary to dynamically adjust the allocation ratio of the new energy power generation to ensure the balance of supply and demand of the system, while giving priority to the use of new energy with lower costs.

[0143] Subsequently, the charging and discharging behavior of the energy storage device is specifically controlled according to the energy storage device scheduling plan. The charging operation of the energy storage device should be carried out when the power of new energy generation is sufficient, and the charging speed and duration should be optimized in combination with the charging efficiency and remaining capacity of the energy storage device to avoid equipment loss or efficiency reduction due to overcharging. When discharge is required, the energy storage device should dynamically adjust the discharge power in combination with the current production load demand and the purchased electricity price, give priority to meeting the load demand of the production task, and avoid purchasing electricity during the peak period of the purchased electricity price. The charging and discharging process also needs to be optimized according to the rated power range and discharge efficiency of the equipment to ensure the safe operation and long-term reliability of the equipment.

[0144] Finally, according to the production task scheduling scheme, the execution order and time window of the production tasks are adjusted. For rigid loads (tasks that must be completed as planned), it is necessary to ensure that their load requirements are met first within the specified time window. For flexible loads (tasks with adjustable execution time), their execution time and operating rate should be optimized according to the dynamic changes of renewable energy power generation and energy storage status. Specifically, during the period of sufficient renewable energy power generation, the execution priority of flexible loads should be appropriately increased to make full use of renewable energy and reduce dependence on purchased electricity. During the period of peak production task load or insufficient renewable energy, the pressure on the overall load of the microgrid can be reduced by reducing the operating rate of the flexible load or postponing its execution time, thereby improving the operating stability of the system.

[0145] During the execution process, it is necessary to continuously adjust the proportion of new energy allocation, the charging and discharging behavior of energy storage equipment, and the production task scheduling strategy in combination with the real-time monitoring data and dynamic changes of the microgrid to cope with sudden energy fluctuations or changes in load demand. In this way, step S103 realizes the coordinated optimization of new energy power generation, energy storage equipment operation, and production tasks, maximizes the utilization efficiency of new energy, reduces the comprehensive energy cost, and ensures the supply and demand balance and operation stability of the microgrid.

[0146] A second embodiment of the present application provides an electronic device, the electronic device comprising:

[0147] processor;

[0148] The memory is used to store a program. When the program is read and executed by the processor, it executes a method for collaborative optimization of microgrids in an industrial park with a high proportion of new energy vehicles provided in the first embodiment of the present application.

[0149] The third embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, a method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park provided in the first embodiment of the present application is executed.

[0150] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park, characterized in that: include: Obtain real-time monitoring data on renewable energy power generation, energy storage equipment status, and production equipment load demand in the microgrid, and use a time series prediction model in combination with historical data to predict renewable energy power generation, available charge and discharge capacity of energy storage equipment, and production load demand in future time periods; Based on the predicted new energy power generation, available charge and discharge capacity of energy storage equipment and production load demand, a mixed integer linear programming model for collaborative optimization with the goal of minimizing comprehensive energy cost is constructed; the mixed integer linear programming model is solved to obtain a new energy power allocation plan, an energy storage equipment scheduling plan and a production task scheduling plan; According to the new energy power allocation plan, the allocation ratio of new energy power between production load and energy storage equipment is adjusted; according to the energy storage equipment scheduling plan, the charging and discharging behavior of the energy storage equipment is controlled; according to the production task scheduling plan, the execution order and time window of the production tasks are adjusted, and the time and rate of the flexible load are optimized.

2. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 1 is characterized in that: The time series prediction model includes an input layer, a feature extraction layer, an attention mechanism layer and an output layer; The input layer is used to receive new energy power generation history data, energy storage equipment status data, production load demand data and environmental impact data, and normalize the received data to obtain a multidimensional time series tensor; The feature extraction layer is used to receive the multidimensional time series tensor provided by the input layer, extract time series features from the received data through stacked bidirectional LSTM units, and obtain a high-dimensional time series feature vector; the stacked bidirectional LSTM units include a first layer of bidirectional LSTM, a second layer of bidirectional LSTM, and a third layer of bidirectional LSTM connected in sequence; wherein the first layer of bidirectional LSTM focuses on the characteristics of new energy power generation and captures its periodic fluctuations; the second layer of bidirectional LSTM focuses on the dynamic characteristics of energy storage equipment; the third layer of bidirectional LSTM processes production load requirements and extracts the periodic characteristics of task loads; The attention mechanism layer is used to receive the high-dimensional time series feature vector output by the feature extraction layer, dynamically adjust the weight of the time step based on the time attention mechanism, and generate a weighted feature vector; wherein, the time attention mechanism specifically includes that for the characteristics of new energy power generation, the attention mechanism tends to highlight the impact of recent meteorological conditions; for the dynamics of energy storage equipment, the weight is mainly allocated to the historical period when the load charge changes drastically; for production load requirements, the attention mechanism emphasizes the key time steps related to the specific task period; The output layer is used to receive the weighted feature vector output by the attention mechanism layer; the weighted feature vector is mapped using a fully connected layer to convert the high-dimensional feature vector into a prediction result, which includes a new energy power generation power prediction, an energy storage device status prediction, and a production load demand prediction; the output layer is implemented using a multi-task learning framework.

3. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 2 is characterized in that: The input layer is specifically used for: Receive new energy power generation history data, energy storage equipment status data, production load demand data and environmental impact data, and format the received data into a unified timestamp sequence; The formatted data is normalized, where the historical data of renewable energy power generation is based on the maximum and minimum power in the past year, the status data of energy storage equipment is based on the rated capacity and maximum power of the equipment, the production load demand data is based on the maximum load demand of the equipment, and the environmental impact data is normalized based on the historical meteorological data range; The normalized time series is divided into time segments of fixed length through the sliding window mechanism, and a multidimensional time series tensor is generated as the input of the feature extraction layer.

4. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 2 is characterized in that: The first-layer bidirectional LSTM is specifically used for: Receive the normalized renewable energy power generation data and meteorological condition data, and process the received data through the bidirectional LSTM unit to capture the periodic fluctuation characteristics of renewable energy power generation; the forward LSTM unit is used to capture the increasing trend of renewable energy power generation characteristics over time, and the reverse LSTM unit captures the degradation pattern of historical data; combine the outputs in the positive and negative directions to generate a complete renewable energy power generation time series feature vector as the input of the second layer of bidirectional LSTM.

5. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 4 is characterized in that: The second-layer bidirectional LSTM is specifically used for: The new energy power generation feature vector from the first layer of bidirectional LSTM is received, and combined with the historical SOC data, charging and discharging power data, and charging and discharging efficiency data of the energy storage device, the input data is processed through the bidirectional LSTM unit to capture the charging and discharging behavior rules and SOC change patterns in the dynamic characteristics of the energy storage device; among them, the forward LSTM unit focuses on the trend prediction of the future charging and discharging behavior of the energy storage device, and the reverse LSTM unit captures the historical charging and discharging abnormal changes, and generates the time series feature vector of the dynamic characteristics of the energy storage device through the fusion of features in the positive and negative directions as the input of the third layer of bidirectional LSTM.

6. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 5 is characterized in that: The third-layer bidirectional LSTM is specifically used for: The dynamic feature vector of the energy storage device is received from the second-layer bidirectional LSTM, and combined with the historical curve of the production load demand in the industrial park, the task scheduling plan, and the task completion time, the input data is processed through the bidirectional LSTM unit to capture the periodic variation characteristics of the production load demand; among them, the forward LSTM unit analyzes the gradual growth pattern of the load demand with the start and end of the task, and the reverse LSTM unit focuses on the sudden load and peak load between historical tasks, and merges the results in the forward and reverse directions into a high-dimensional time series feature vector of the comprehensive production load demand as the input of the attention mechanism layer.

7. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 2 is characterized in that: The output layer is specifically used for: Receive the weighted feature vectors from the attention mechanism layer, and assign the weighted feature vectors to different task prediction branches according to the characteristics of renewable energy power generation, the dynamic characteristics of energy storage equipment, and the characteristics of production load demand; In the renewable energy power generation prediction branch, weighted features related to meteorological conditions are used to further process environmental data including light intensity and wind speed, and the renewable energy power generation prediction results for the future time period are output; In the energy storage device state prediction branch, based on the historical load state, charge and discharge efficiency and dynamic characteristic weighted features of the energy storage device, combined with the impact of future task scheduling load on the charge and discharge behavior of the energy storage device, the available charge and discharge capacity and state of the energy storage device in the future time period are output; In the production load demand forecasting branch, the weighted features related to the production task period are used, combined with the historical task start and end time and load trends, to generate a production task load demand forecast for the future time period; Through the multi-task learning framework, the underlying features shared by the above prediction branches are separated from the specific features of each branch, and the joint loss function is used to optimize the prediction accuracy of each task, ensuring that the prediction results of renewable energy power generation, energy storage equipment status and production load demand are coordinated with each other to meet the needs of microgrid collaborative optimization.

8. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 2 is characterized in that: The attention mechanism layer is specifically used for: Receive the high-dimensional time series feature vector output by the feature extraction layer, and divide it into feature sub-vectors corresponding to the new energy power generation characteristics, the dynamic characteristics of the energy storage equipment, and the production load demand characteristics; For the characteristics of renewable energy power generation, the correlation between each time step and the current time is calculated based on the historical change trend of meteorological conditions, and the time steps with significant changes in light intensity and wind speed conditions are highlighted, and their impact on renewable energy power generation prediction is emphasized by assigning higher weights; For the dynamic characteristics of energy storage equipment, analyze the historical load state changes of energy storage equipment and the periods of drastic changes in charging and discharging power, identify the time steps that are closely related to the future energy storage state, assign higher weights to the identified time steps, and enhance the accuracy of the model's prediction of the energy storage equipment state; For production load demand characteristics, based on the scheduling plan of production tasks, key time steps close to the task start time or end time are identified and assigned higher weights, so that the characteristics of these time steps play a more important role in load demand forecasting; The weight assigned to each time step is weighted with the corresponding characteristic sub-vector to generate weighted characteristic vectors of various characteristics; the weighted characteristic vectors of new energy power generation, energy storage dynamics and production load demand are fused into weighted characteristic vectors.

9. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 1, characterized in that: The objective function of the mixed integer linear programming model is shown in the following formula (1): Among them, C total represents the total cost; C grid represents the cost of purchased electricity; C storage Represents the operating cost of energy storage equipment; C task represents the production task scheduling cost; C renewable represents the penalty cost for not using new energy; the cost of purchasing electricity C grid It is expressed using the following formula (2): λ1, λ2, λ3 and λ4 are weight coefficients; ∈ is a small positive number used to avoid the denominator being 0; Where T represents the total number of discrete time steps in the optimization period; P grid,t represents the purchased power at time t; λ grid represents the unit cost of purchased electricity; The energy storage equipment operating cost C storage Use the following formula (3) to express: Where T represents the total number of discrete time steps in the optimization period; P charge,t represents the charging power of the energy storage device at time t; P discharge,t represents the discharge power of the energy storage device at time t; η charge Represents the charging efficiency of the energy storage device; η discharge Represents the discharge efficiency of the energy storage device; storage Indicates the unit energy storage equipment operating cost; Among them, the production task scheduling cost C task The following formula (4) is used to express: Where T represents the total number of discrete time steps in the optimization period; N represents the total number of production tasks; L i,t represents the load demand of production task i at time t; α i represents the priority weight of task i; represents the actual scheduling time of task i at time t; represents the optimal scheduling time of task i at time t; λ task represents the unit dispatch cost weight; The penalty cost C for not utilizing the new energy renewable Use the following formula (5) to express: Where T represents the total number of discrete time steps in the optimization period; The maximum power of renewable energy generation at time t; Indicates the actual power used by renewable energy generation at time t; renewable It indicates the penalty cost weight for the unit not utilizing new energy.

10. The method for collaborative optimization of microgrids in a high-proportion new energy vehicle industrial park according to claim 1, characterized in that: The constraints of the mixed integer linear programming model include: (a) Power balance constraint, expressed as follows: Among them, P gen,t Represents the renewable energy power generation at time t; P grid,t represents the purchased power at time t; P discharge,t Represents the energy storage discharge power at time t; P load,t represents the production load demand at time t; P charge,t represents the energy storage charging power at time t; T represents the total number of discrete time steps in the optimization time period; (b) Energy storage device state constraint, expressed using the following formula (7): Among them, SOC t+1 Indicates the state of charge of the energy storage device at time t+1; SOC t represents the state of charge of the energy storage device at time t; Δt is the time step; η charge Represents the charging efficiency of the energy storage device; P charge,t represents the charging power of the energy storage device at time t; P discharge,t represents the discharge power of the energy storage device at time t; η discharge represents the discharge efficiency of the energy storage device; T represents the total number of discrete time steps in the optimization period; (c) The operating range constraints of the energy storage device are expressed using the following formulas (8) and (9): Among them, SOC t SOC is the state of charge of the energy storage device at time t, indicating the current power level of the energy storage device; min SOC is the minimum allowable state of charge of the energy storage device, indicating the minimum power limit of the energy storage device when discharging; min P is the highest state of charge of the energy storage device, indicating the maximum power limit of the energy storage device when charging; charge,t represents the charging power of the energy storage device at time t; Indicates the maximum charging power of the energy storage device; P discharge,t represents the discharge power of the energy storage device at time t; Indicates the maximum discharge power of the energy storage device; (d) The constraint condition of new energy power fluctuation is expressed by the following formula (10): Among them, P gen,t is the renewable energy power generation at time t; is the maximum power generation of renewable energy at time t; P gen,t+1 is the renewable energy power generation at time t+1; R ramp is the power ramp limit; (e) The production task scheduling constraints are expressed using the following formulas (11) and (12): Among them, T i,start Indicates the start time of task i; Indicates the earliest start time of task i; represents the latest start time of task i; L i,t represents the load requirement of task i at time i; L i,max represents the maximum load requirement of task i; T i,start represents the actual start time of task i; T i,end represents the actual end time of task i; T represents the total number of discrete time steps in the optimization period; N represents the total number of tasks; (f) The external power purchase restriction constraint is expressed using the following formula (13): Among them, P grid,t is the power purchased from the external grid at time t; is the maximum power supply of the external power grid; T represents the total number of discrete time steps in the optimization period; (g) The renewable energy priority consumption constraint is expressed using the following formulas (14) and (15): Among them, P load,t represents the production load demand at time t; P charge,t represents the charging power of the energy storage device at time t; P gen,t is the renewable energy power generation at time t; P grid,t is the power purchased from the external grid at time t; It represents the actual power used by renewable energy generation at time t; T represents the total number of discrete time steps in the optimization period.

Images