A power grid-energy storage collaborative scheduling method, device and medium

Abstract

The present application relates to a kind of power grid-energy storage collaborative scheduling method, equipment and medium, the method includes: collecting multi-source heterogeneous data and pre-processing to obtain standardized data;Standardized data is input into the LSTM network embedded in feature attention mechanism, predicts power generation;With the minimum of micro-grid operating cost as target, set constraint, build day-ahead-intra-day-real-time three-layer bidirectional coupling collaborative scheduling model;Collect real-time features to build multidimensional state vector, adaptively select convex optimization solver to solve model, get scheduling plan power;Based on illumination fluctuation, load rate and energy storage safety margin, calculate dynamic threshold, when the deviation of actual execution power and planned power exceeds threshold, reverse correct LSTM network parameters, and re-execute the prediction to scheduling step, get updated scheduling plan.The method realizes the prediction drift inhibition, scheduling mismatch real-time correction and extreme condition robust operation, significantly improves the system economy, security and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, specifically to a grid-energy storage coordinated dispatch method, equipment, and medium. Background Technology

[0002] With the rapid development of renewable energy, microgrids, as an effective carrier for integrating distributed energy resources, play a crucial role in improving energy efficiency and promoting low-carbon transformation. New energy generating units, as the core component of microgrids, exhibit strong randomness and volatility in power output due to factors such as wind, sunlight, and temperature, posing challenges to the stable operation and economic dispatch of microgrids. Achieving coordinated optimization between the power grid and energy storage systems based on a "prediction-dispatch" framework has become the mainstream technical approach for microgrids to mitigate renewable energy fluctuations and improve energy utilization.

[0003] Currently, patent application publication number CN119513544A discloses a method and system for accurately predicting the power generation of photovoltaic power stations in the next time period by dividing the photovoltaic power stations into regions, constructing multiple power generation impact data sequences, and combining them with the actual power generation of the current time period. Patent application publication number CN118249420B discloses a method for adjusting the current operating power of each system based on the prediction results of various models, including hourly predictions of building power load, photovoltaic system power generation, battery discharge power, air conditioning system power consumption, and other power consumption. Patent application publication number CN115842375A discloses a method for determining the optimal system scheduling scheme by dividing traditional day-ahead scheduling into different time scales based on the power equipment regulation characteristics, combining power system DC power flow models and wind-solar load prediction models, and considering line transmission constraints, energy storage, and remaining energy in energy storage. Patent application publication number CN116488180B discloses a method for determining the operating status by predicting the power and expected generation of new energy power plants, calculating the planned changes and change coefficients for each power plant, and adjusting the plans of each new energy power plant according to the operating status while aiming to maximize grid stability. Patent application publication number CN118263893A discloses a device and system for calculating the load demand of a power plant area based on energy dispatch instructions, and adjusting the dispatch control strategy according to the supply-demand relationship between the predicted power values ​​of wind, solar, energy storage, and charging equipment in the area and the load demand.

[0004] However, current related technologies generally adopt a serial implementation approach of "data acquisition - power prediction - optimized scheduling". Although this improves the rationality of scheduling to a certain extent, it still has the following common technical defects: (1) Prediction and scheduling are disconnected and do not form a true closed loop. Existing methods mostly use power prediction as the static input of scheduling. The prediction model only uses its own error as the optimization target, and the scheduling execution result cannot correct the prediction model parameters in reverse. In extreme scenarios such as continuous cloudy days and sudden load changes, the prediction deviation continues to accumulate, causing the scheduling strategy to gradually deviate from the optimal, making it difficult to achieve long-term stable operation.

[0005] (2) The prediction deviation lacks a root cause correction mechanism. Most schemes only monitor and evaluate the prediction error, and cannot trace the systematic deviation of the prediction model back to the source through the scheduling deviation. The prediction model does not have the ability to self-optimize and self-evolve with the operating scenario, and it is difficult to cope with complex and ever-changing power grid conditions.

[0006] (3) The scheduling architecture is unidirectional and lacks robustness. Existing multi-timescale scheduling mostly adopts the unidirectional transmission mode of "day-to-day-real-time", lacking the upward feedback channel between the real-time layer and the intra-day layer, and between the intra-day layer and the day-to-day layer. It cannot dynamically correct the upper-level scheduling plan according to the actual operation deviation, and the whole exhibits open-loop control characteristics with weak anti-disturbance capability.

[0007] (4) Fixed and rigid optimization strategies. Convex optimization solvers often adopt a fixed priority trial mechanism, without dynamically selecting the optimal solver based on the problem size, the degree of ill-conditioned constraints, and real-time requirements. This results in low solution efficiency, poor numerical stability, and difficulty in meeting the real-time scheduling needs of microgrids.

[0008] (5) The deviation trigger threshold is a fixed value, which has poor adaptability to different scenarios. The trigger conditions of the closed loop of the "prediction-scheduling" framework mostly adopt fixed thresholds, without comprehensively considering factors such as wind force level, light intensity fluctuation, load level, and energy storage SOC safety margin. This can easily lead to false triggering under stable operating conditions and delayed response under extreme operating conditions, affecting system operating efficiency and equipment safety.

[0009] Therefore, there is an urgent need for a grid-energy storage coordinated dispatch method to solve the problems existing in the above-mentioned technologies. Summary of the Invention

[0010] This invention provides a grid-energy storage collaborative scheduling method, device, and medium, which aims to solve the technical problems in existing grid-energy storage collaborative scheduling methods, such as the separation of prediction and scheduling, lack of scheduling deviation reverse correction mechanism, unidirectional open-loop scheduling architecture, fixed and rigid solution strategy, and poor adaptability to deviation trigger threshold scenarios.

[0011] To achieve the above objectives, the first aspect of the present invention provides a grid-energy storage coordinated dispatch method, comprising the following steps: Step S100: Collect multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain standardized data; Step S200: Input the standardized data into a long short-term memory network with an embedded feature attention mechanism to predict the power generation of the new energy unit and obtain the predicted power. Step S300: With the goal of minimizing the total operating cost of the microgrid, set constraints, construct a three-layer bidirectional coupled collaborative scheduling model of day-ahead, intraday, and real-time, and input the predicted power into the collaborative scheduling model; Step S400: Collect real-time operating characteristics of the system to construct a multi-dimensional state vector, and adaptively select a convex optimization solver based on the multi-dimensional state vector to solve the cooperative scheduling model and obtain the scheduling plan power; Step S500: Calculate the scenario adaptive dynamic threshold based on the standard deviation of light intensity fluctuation, load factor and energy storage safety margin, obtain the actual executed power and compare it with the scheduling plan power to obtain the deviation. When the absolute value of the deviation exceeds the scenario adaptive dynamic threshold, correct the network parameters of the long short-term memory network in reverse according to the deviation, and re-execute the step of predicting the power generation of new energy units based on the corrected network parameters until the step of obtaining the scheduling plan power is obtained, and obtain the updated scheduling plan power.

[0012] Furthermore, the collection of multi-source heterogeneous data includes: collecting historical power generation, load power, meteorological environment information, and energy storage data of new energy units as multi-source heterogeneous data. The meteorological environment information includes temperature, humidity, and light intensity, and the energy storage data includes battery state of charge, upper and lower limits of charging and discharging power, rated power of diesel engines, time-of-use electricity price of the power grid, and power grid operation status. The method for preprocessing the multi-source heterogeneous data includes: outlier removal, missing value imputation, and normalization of the multi-source heterogeneous data.

[0013] Furthermore, the method for inputting the standardized data into a long short-term memory network with an embedded feature attention mechanism to predict the power generation of new energy units includes: Construct a Long Short-Term Memory (LSTM) network with an embedded feature attention mechanism; wherein the LSM network includes a forget gate, an input gate, an output gate, and a memory unit, and the feature attention mechanism is used to dynamically calculate attention weight coefficients for each input feature and to use the attention weight coefficients to perform weighted updates on the standardized data; The Long Short-Term Memory (LSTM) network with embedded feature attention mechanism is trained offline using preprocessed historical multi-source heterogeneous data. The mean squared error is used as the loss function, and the LSM network parameters are updated through backpropagation until the loss function converges, thus obtaining the trained LSM network. The data, which is collected in real time and standardized, is input into the trained Long Short-Term Memory network. After forward propagation calculation through the forget gate, input gate, output gate and memory unit, the predicted value of the power generation of the new energy unit is output as the predicted power.

[0014] Furthermore, with the goal of minimizing the total operating cost of microgrids, the method for constructing a three-layer bidirectional coupled collaborative scheduling model (day-ahead, intraday, and real-time) with defined constraints includes: Construct an objective function, which aims to minimize the total operating cost of the microgrid, and its expression is:

[0015] in, for Power grid interaction at all times for Time-of-use electricity pricing at any given moment for The actual output of the diesel engine at any given time. This is the operating cost coefficient per unit power of the diesel engine; This represents the total number of scheduling periods; The constraints are defined, including energy balance constraints, energy storage battery state constraints, and device power constraints; the energy balance constraints are as follows:

[0016] in, for Real-time new energy power forecast and They are respectively Real-time energy storage discharge power and charging power, for Real-time load power; The energy storage battery state constraints include the dynamic update equation for the state of charge and upper and lower bound constraints as follows:

[0017] in, for Energy storage status at all times For battery capacity, For time intervals, Indicates the charging and discharging efficiency of the energy storage system; The power constraints of the equipment include upper and lower limits of charging and discharging power of the energy storage system and upper and lower limits of generating power of the diesel engine.

[0018] in, This represents the lower limit of the actual output power of the diesel engine. This represents the upper limit of the actual output power of the diesel engine; This represents the lower limit of the discharge power of the energy storage system. This represents the upper limit of the discharge power of the energy storage system. The lower limit of the charging power for energy storage systems; The upper limit of the charging power of the energy storage system; A three-layer bidirectional coupled scheduling architecture is constructed, consisting of a day-ahead layer, an intraday layer, and a real-time layer. The day-ahead layer generates a baseline scheduling plan with a first period and a first step length, and transmits the baseline scheduling plan and prediction confidence interval to the intraday layer. The intraday layer receives deviation information uploaded by the real-time layer with a second period as the rolling window and a second step length, dynamically adjusts the scheduling trajectory, and sends a re-optimization command to the day-ahead layer when the accumulated deviation exceeds a preset threshold. The real-time layer outputs an instantaneous power correction command with a third period and a third step length, and feeds back the real-time deviation characteristics to the intraday layer.

[0019] Furthermore, the method for constructing a multi-dimensional state vector based on the real-time operating features of the acquisition system, and adaptively selecting a convex optimization solver based on the multi-dimensional state vector, includes: The system collects real-time operational characteristics and constructs a multi-dimensional state vector, expressed as follows:

[0020] in, It is a multidimensional state vector. for Number of decision variables at any given time for Number of time-bound constraints for The condition number of the constraint matrix at any given time. for Time-varying nonlinear relaxation residuals for The remaining time of the current scheduling cycle; The multidimensional state vector is input into a lightweight decision tree model trained offline, and the online performance of multiple convex optimization solvers is evaluated. The optimal solver is selected based on the evaluation results.

[0021] Furthermore, the method for solving the cooperative scheduling model to obtain the scheduling plan power includes: applying the selected convex optimization solver to the cooperative scheduling model; if the solution returns an optimal solution, then outputting the optimal solution as the scheduling plan power; if the solution is abnormal, then switching to a preset backup convex optimization solver to solve the cooperative scheduling model again, and using the solution result of the backup convex optimization solver as the scheduling plan power.

[0022] Furthermore, based on the standard deviation of light intensity fluctuation, load factor, and energy storage safety margin, a scenario adaptive dynamic threshold is calculated to obtain the actual executed power and compare it with the scheduling plan power. When the absolute value of the deviation exceeds the scenario adaptive dynamic threshold, the following steps are taken: The scene-adaptive dynamic threshold is calculated using the following expression:

[0023] in, For scene-adaptive dynamic thresholds, Based on the threshold, for Standard deviation of light intensity fluctuation over time for real-time load factor for Always have a safe and sufficient energy storage capacity. It is an adaptive function, which adjusts the threshold value according to the fluctuation of light intensity, load level and energy storage status; Real-time acquisition of actual execution power Obtain the scheduling plan power Calculate the absolute value of the deviation The deviation is compared with the scene adaptive dynamic threshold. If the absolute value of the deviation is greater than the scene adaptive dynamic threshold, the closed-loop correction process is triggered.

[0024] Furthermore, the method for obtaining the updated scheduling plan power involves correcting the network parameters of the long short-term memory network in reverse based on the deviation, and re-executing the step of predicting the power generation of the new energy units based on the corrected network parameters until the scheduling plan power is obtained. Based on the aforementioned deviation, a new energy dispatch output deviation loss function is constructed:

[0025] in, For the power output deviation loss function of new energy dispatch, These are time-varying weighting coefficients; when the energy storage state of charge approaches the safety boundary, Automatically increases in size; Using the new energy dispatch output deviation loss function, the weight matrix and feature attention weight coefficients of the long short-term memory network are corrected in reverse through an adaptive gradient descent mechanism to obtain the corrected long short-term memory network. Based on the modified Long Short-Term Memory network, the step of predicting the power generation of new energy units is re-executed to obtain the updated predicted power. Then, based on the updated predicted power, the steps of constructing the collaborative scheduling model and solving it are re-executed until the updated scheduling plan power is obtained.

[0026] To achieve the above objectives, a second aspect of the present invention provides an electronic device including a memory and a processor, the memory being used to store a program that supports the processor in executing the grid-energy storage coordinated scheduling method, and the processor being configured to execute the program stored in the memory.

[0027] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the grid-energy storage coordinated scheduling method.

[0028] The beneficial effects of this invention are: Compared with existing technologies, the present invention provides a grid-energy storage collaborative scheduling method, device, and medium. By constructing a prediction-scheduling bidirectional coupled closed-loop architecture, the scheduling execution deviation drives the network parameters of the long short-term memory network to adaptively and iteratively update, thereby suppressing prediction drift. A three-layer bidirectional coupled scheduling architecture (day-ahead, intraday, real-time) introduces uplink feedback channels from the real-time layer to the intraday layer and from the intraday layer to the day-ahead layer, breaking the traditional unidirectional rolling mode to improve robustness. By collecting real-time operating characteristics of the system to construct a multi-dimensional state vector and adaptively selecting a convex optimization solver, the inefficiency and numerical instability caused by fixed solution strategies are avoided. By calculating adaptive dynamic thresholds based on the standard deviation of light intensity fluctuations, load factor, and energy storage safety margin, the problem of false triggering under stable conditions and delayed response under extreme conditions with fixed thresholds is overcome. Thus, the invention achieves comprehensive collaborative scheduling with deep coupling of prediction and scheduling, root cause correction of deviations, bidirectional feedback across multiple time scales, intelligent adaptation of solution strategies, and adaptive threshold scenarios. This effectively solves the technical problems of existing technologies, such as the disconnect between prediction and scheduling, lack of reverse correction, unidirectional open-loop architecture, rigid solution strategies, and poor adaptability to threshold scenarios. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0030] Figure 1 This is a flowchart of a grid-energy storage coordinated scheduling method disclosed in an embodiment of the present invention.

[0031] Figure 2 This is a structural diagram of a feature attention mechanism design disclosed in an embodiment of the present invention.

[0032] Figure 3 This is a structural diagram of a multi-dimensional attention-enhanced LSTM new energy power output timing intelligent prediction module disclosed in an embodiment of the present invention.

[0033] Figure 4This is a comparison chart of errors of different models in the photovoltaic power generation prediction task disclosed in an embodiment of the present invention, wherein... Figure 4 (a) in the figure is a comparison chart of root mean square error (RMSE). Figure 4 (b) in the figure is a comparison chart of mean absolute error (MAE). Figure 4 (c) in the figure is a comparison chart of the coefficient of determination (R²).

[0034] Figure 5 This is a comparison chart of the fitting effects of different models on the photovoltaic power generation prediction task, as disclosed in an embodiment of the present invention, between the predicted curve and the actual value curve.

[0035] Figure 6 This is an output diagram of a 24-hour dispatch strategy model for a microgrid disclosed in an embodiment of the present invention, wherein... Figure 6 (a) in the graph shows the 24-hour electricity price of the microgrid as a function of time. Figure 6 (b) in the figure is a graph showing the power of each device over time in the 24-hour dispatch strategy of the microgrid.

[0036] Figure 7 This is a flowchart of a prediction-scheduling bidirectional coupled closed loop and scheduling error reverse correction disclosed in an embodiment of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0038] According to embodiments of the present invention, it should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the following methods, in some cases the steps shown or described may be executed in a different order than that shown here.

[0039] like Figure 1 As shown, this invention provides a grid-energy storage coordinated dispatch method, which includes: Step S100: Collect multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain standardized data; The data collected in this embodiment includes historical power generation, load power, meteorological environmental information, and energy storage system-related data. Meteorological environmental information includes temperature, humidity, and light intensity; energy storage system-related data includes battery state of charge, upper and lower limits of charge / discharge power, diesel engine rated power, grid time-of-use pricing, and grid operating status. The above data originates from different data acquisition terminals or monitoring systems, and the sampling frequencies, data formats, and units may vary. Therefore, preprocessing is required to create a unified and standardized input dataset.

[0040] Secondly, outlier removal is performed on the collected raw data. This embodiment uses the 3σ criterion to identify and remove outliers in the data to avoid interference from outliers in subsequent prediction and scheduling.

[0041] Then, missing data resulting from outlier removal is filled in. This embodiment uses nearest-neighbor interpolation to handle missing values, thereby ensuring the continuity and integrity of the data sequence.

[0042] Finally, to eliminate numerical deviations caused by differences in units between different data sources, the processed data is normalized. This embodiment uses the Min-Max normalization method to linearly map the original data to the interval [0, 1]. The normalization transformation rules are as follows:

[0043] in, This represents the normalized data. The original data to be normalized. This represents the minimum value in the original data sample. This represents the maximum value in the original data sample. After the above preprocessing, standardized data is obtained.

[0044] Step S200: Input the standardized data into a long short-term memory network with an embedded feature attention mechanism to predict the power generation of the new energy unit and obtain the predicted power. In this embodiment, considering the advantages and wide application of Long Short-Term Memory (LSTM) networks in solving time-series prediction tasks, LSTM is selected as the basic model of this invention. The LSTM network constructed in this embodiment mainly consists of a forget gate, an input gate, an output gate, and a memory unit. At each time step... In this process, the LSTM network selectively forgets, updates, and outputs information based on the output of the previous time step and the input of the current time step through a gating mechanism, and finally obtains the predicted value of the current time step.

[0045] Specifically, the forget gate determines how much information from the previous memory unit's state needs to be retained in the current time step. The output calculation formula for the forget gate is as follows:

[0046] in, The output value of the forget gate. This represents the Sigmoid activation function, with output values ​​between 0 and 1. Here is the weight matrix for the forget gate; This represents the network output status from the previous moment. This represents the current input state. This is the bias vector for the forget gate.

[0047] The input gate controls how much of the current input information needs to be stored in memory. The calculation of the input gate involves two parts: the input gate coefficient and the candidate memory cell state, as shown in the following formula:

[0048]

[0049] in, This represents the output value of the input gate (i.e., the input gate coefficient). Candidate memory cell state; and These are the weight matrices for the input gate and candidate memory units, respectively; and For the corresponding bias vector, It is the hyperbolic tangent activation function.

[0050] The output gate determines the network's output value based on the current state of the memory cells. The formulas for calculating the output gate coefficients and the network output at the current moment are as follows:

[0051]

[0052] in, This represents the output value of the output gate at the current moment (i.e., the output gate coefficient). This is the network output at the current moment; This is the weight matrix of the output gate. This is the bias vector for the output gate. This represents the current state of the memory unit.

[0053] The state update of the memory unit is jointly controlled by the forget gate and the input gate. The specific update method is as follows:

[0054] in, The state of the memory unit at the current moment. This represents point-by-point product. This represents the state of the memory unit from the previous moment.

[0055] To enhance the LSTM network's ability to extract key meteorological features, this embodiment introduces a feature attention mechanism into the LSTM network's input layer. This attention mechanism dynamically assigns different weight coefficients to each input feature, making the model pay more attention to factors that significantly impact the prediction of new energy power generation (such as light intensity and temperature), thereby improving prediction accuracy. Specifically, at each time step... The current input state The attention weight vector is used for weighted updates. First, the attention weight vector is calculated through a neural network layer. :

[0056] in, Here is the weight matrix for the attention mechanism. Here, is the bias vector, and ReLU is the activation function.

[0057] Then, the attention weight vector is normalized to obtain the first... Attention weight coefficients for each feature Specifically, such as Figure 2 The diagram shown is a structural design of the feature attention mechanism of this invention. The attention weight coefficients for each feature are calculated using the following Softmax normalization process:

[0058] in, For the first Attention weight vector components of each feature, This represents the total number of input features. Through this normalization process, the sum of the attention weight coefficients of all features is 1, thus achieving a probabilistic measure of the importance of different features.

[0059] The weighted updated input data is represented as follows:

[0060] in, Indicates time The input data is weighted and updated using a feature attention mechanism. The attention weight coefficients for temperature features. These are the actual values ​​of the temperature characteristics. The attention weight coefficient for wind speed characteristics. This represents the actual value of the wind speed characteristic.

[0061] After completing the network architecture design, the LSTM network with embedded feature attention mechanism needs to be trained offline. For example... Figure 3 The diagram shown illustrates the structure of the multi-dimensional attention-enhanced LSTM-based intelligent prediction module for renewable energy output time series, constructed according to this invention. This module cascades a feature attention mechanism with an LSTM time series network to achieve adaptive weighting of input features and time-dependent modeling.

[0062] This embodiment uses preprocessed historical multi-source heterogeneous data as training samples and employs mean squared error (MSE) as the loss function, the calculation formula of which is as follows:

[0063] in, The value of the loss function. The total number of training samples, For the first The actual power generation value of each sample This represents the predicted power generation value from the model. During backpropagation, the gradient of each network parameter is calculated based on the loss function, and the weight matrix and bias vector are updated by the optimizer. The training process is repeated until the loss function converges to a small value.

[0064] To scientifically evaluate the accuracy of the prediction model, this embodiment selects the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (COP). RMSE is used as an evaluation indicator. The formula for calculating RMSE is:

[0065] RMSE retains the dimensions of the original data and reflects the degree of deviation between the predicted and actual values; the smaller the value, the more accurate the prediction. The formula for MAE is:

[0066] MAE reflects the average level of the absolute error between the predicted and actual values; the smaller the value, the better the robustness of the model. The calculation formula is:

[0067] in, This is the mean of the true values. It measures the model's ability to explain data variation; the closer its value is to 1, the higher the model's goodness of fit.

[0068] After offline training is completed, standardized data collected in real time and subjected to the same preprocessing operations are input into the trained LSTM network with embedded feature attention mechanism. The network calculates and outputs the predicted power generation value of the new energy unit through forward propagation, thus obtaining the predicted power. This is for use by subsequent scheduling models.

[0069] like Figure 4 , Figure 5 The diagram shows a comparison of the proposed model (FALSTM) with other comparative models (XGBoost, LSTM, SVR, and MLP) in the task of predicting photovoltaic power generation. It is evident that the model's prediction curve best matches the actual power curve.

[0070] Step S300: With the goal of minimizing the total operating cost of the microgrid, set constraints, construct a three-layer bidirectional coupled collaborative scheduling model of day-ahead, intraday, and real-time, and input the predicted power into the collaborative scheduling model; In this embodiment, the construction of the collaborative scheduling model in this step mainly consists of three processes: 1. Construction of the scheduling objective function; 2. Setting of multi-dimensional constraints; 3. Design of a three-layer bidirectional coupled scheduling architecture. Its specific implementation process mainly includes: First, the objective function for microgrid coordinated scheduling is constructed. This embodiment prioritizes minimizing the total operating cost of the microgrid, comprehensively considering both grid power purchase costs and diesel engine operating costs. The expression for the objective function is as follows:

[0071] in, for The power exchange between the grid and the power source at any given time is represented by a positive value indicating that electricity is purchased from the grid, and a negative value indicating that electricity is sold to the grid. for The time-of-use electricity price is provided by the electricity market or the grid company. for The actual output of the diesel engine at any given time. This is the unit power operating cost coefficient for diesel engines, reflecting the fuel consumption and maintenance costs of diesel engine power generation; This represents the total number of scheduling periods.

[0072] By minimizing the above objective function, the microgrid can minimize electricity purchase expenditure and diesel generator generation costs while meeting load demand, thereby improving its operational economy.

[0073] Secondly, multi-dimensional constraints are set for the collaborative scheduling model. This embodiment mainly considers the following three types of constraints: (1) Energy balance constraint. This constraint ensures that the power supply and demand of the microgrid are balanced at every moment, that is, the sum of the output of all power sources equals the load demand. The expression is:

[0074] in, The result obtained in step S200 Real-time forecast of new energy power (such as photovoltaic or wind power). and They are respectively The energy storage discharge power and charging power at all times (discharge is positive, charging is negative). for Load power at any given time.

[0075] (2) Energy storage battery state constraints. This mainly includes the dynamic update equations for the energy storage state of charge (SOC) and the upper and lower bound constraints for the SOC. The SOC dynamic update equations are:

[0076] in, for Energy storage status at all times For battery capacity, For time intervals, This indicates the charging and discharging efficiency of the energy storage system.

[0077] Meanwhile, the energy storage SOC must meet upper and lower limits: This is to prevent overcharging or over-discharging and protect battery life.

[0078] (3) Equipment power constraints. This mainly includes the upper and lower limits of the charging and discharging power of the energy storage system and the upper and lower limits of the diesel engine power generation.

[0079]

[0080] in, This represents the lower limit of the actual output power of the diesel engine. This represents the upper limit of the actual output power of the diesel engine; This represents the lower limit of the discharge power of the energy storage system. This represents the upper limit of the discharge power of the energy storage system. The lower limit of the charging power for energy storage systems; The upper limit of the charging power of the energy storage system.

[0081] Furthermore, a three-layer bidirectional coupled scheduling architecture is constructed, encompassing day-to-day, intraday, and real-time modes. To overcome the insufficient robustness of the traditional "day-to-day, intraday, and real-time" unidirectional rolling mode, this embodiment designs a bidirectional coupled three-layer scheduling architecture, where downlink reference transmission and uplink deviation feedback exist between each layer.

[0082] The day-ahead dispatch layer operates on a relatively large time scale (e.g., a 24-hour period with a 1-hour step). Based on the predicted renewable energy power output from step S200, this layer solves the objective function under the aforementioned constraints to generate a baseline dispatch plan for the next 24 hours, including grid interaction power, diesel engine output, and energy storage charging and discharging power for each time period. Simultaneously, the day-ahead layer outputs the prediction confidence interval and transmits the baseline dispatch plan as boundary constraints to the intraday dispatch layer.

[0083] The intraday scheduling layer operates on a medium timescale (e.g., a 4-hour rolling window and a 15-minute step). This layer receives the baseline scheduling plan from the day-ahead layer as a reference trajectory, and also receives deviation information uploaded by the real-time layer. Based on the latest forecast data and real-time feedback, the intraday layer performs rolling optimization corrections near the baseline plan, dynamically adjusting the scheduling trajectory. When the accumulated deviation exceeds a preset threshold, the intraday layer sends a re-optimization command to the day-ahead layer, triggering the day-ahead layer to recalculate the scheduling plan for subsequent time periods.

[0084] The real-time scheduling layer operates on a shorter timescale (e.g., a 30-minute cycle with a 5-minute step). This layer is used to quickly respond to instantaneous changes such as sudden changes in photovoltaic power, load fluctuations, and energy storage SOC jumps, outputting immediate power correction commands. The real-time layer feeds back the real-time deviation characteristics generated during execution (such as the difference between actual and planned power, deviation duration, etc.) to the intraday layer, forming a closed-loop correction.

[0085] Through the bidirectional coupling mechanism combining "downlink reference transmission" and "uplink deviation feedback" mentioned above, the three-layer scheduling architecture achieves collaborative adaptive optimization across multiple time scales: the real-time layer's correction commands can respond promptly to instantaneous disturbances, the intraday layer performs rolling optimization based on real-time feedback, and the day-ahead layer replans when necessary, thereby significantly improving the robustness and economy of the microgrid in the face of extreme operating conditions and load mutations.

[0086] Finally, the predicted power obtained in step S200, along with the load data, energy storage status, time-of-use electricity price, and other parameters acquired from the system, are input into the constructed collaborative scheduling model as the basic input for subsequent solutions.

[0087] Step S400: Collect real-time operating characteristics of the system to construct a multi-dimensional state vector, and adaptively select a convex optimization solver based on the multi-dimensional state vector to solve the cooperative scheduling model and obtain the scheduling plan power; In actual microgrid dispatching, different operating conditions place significant differences in the performance requirements of the optimization solver. For example, small-scale problems may require high solution accuracy, while large-scale problems emphasize real-time performance; when the constraint matrix is ​​ill-conditioned, a solver with stronger numerical stability is needed. Traditional methods employ a fixed-priority solver trial mechanism, which struggles to balance solution accuracy, real-time performance, and numerical stability. Therefore, this embodiment proposes a dynamic solver intelligent selection mechanism based on real-time situational awareness.

[0088] First, real-time operating characteristics of the system are collected to construct a multi-dimensional state vector. Before each scheduling calculation begins, the current-moment system operating characteristics are extracted from the microgrid energy management system and combined into a multi-dimensional evaluation vector. Its expression is:

[0089] in, It is a multidimensional state vector. for The number of decision variables at any given time, such as the total number of variables including energy storage charging and discharging power, grid interconnection power, and diesel engine output, reflects the size of the problem. for The number of constraints at any given time, including energy balance constraints, upper and lower limits of energy storage SOC, and equipment power constraints, is also used to assess the complexity of the problem. for The condition number of the constraint matrix is ​​a measure of the ill-conditioned nature of the matrix. A larger condition number indicates that the constraint matrix is ​​closer to singularity, and the more likely numerical instability will occur during the solution process. for The time-varying nonlinear relaxation residuals are used to assess the degree of convexity deviation in real-world scheduling problems, as these residuals may contain nonlinear factors. for The remaining time of the current scheduling cycle is the available time window from the current moment until the scheduling instruction must be output. This parameter reflects the requirements for real-time solution.

[0090] The aforementioned multidimensional state vectors characterize the current scheduling situation from multiple dimensions, including problem size, numerical stability, degree of convexity deviation, and real-time requirements.

[0091] Secondly, online performance evaluation of multiple convex optimization solvers is performed based on the multidimensional state vector, and the optimal solver is adaptively selected. In this embodiment, a lightweight decision tree model is pre-trained offline, and the input of this model is the aforementioned multidimensional state vector. The output is the performance score of each candidate solver under the current operating conditions. Candidate solvers include, but are not limited to, mainstream convex optimization solvers such as CLARABEL, ECOS, and SCS.

[0092] The training process of the decision tree model is as follows: A large number of historical scheduling instances under different operating conditions are collected. Each instance contains a multi-dimensional state vector and labels for the actual solution time, solution accuracy, and numerical stability when using different solvers under that operating condition. The decision tree model is trained through supervised learning, enabling it to predict the overall performance of each solver based on the input state vector.

[0093] During actual scheduling execution, the multi-dimensional state vector constructed at the current moment will be... The trained decision tree model is input, and the model outputs a recommended optimal solver. This embodiment executes the following intelligent matching strategy based on the decision results: If the problem is small in size, the constraint matrix condition number is low (i.e., the ill-conditioned degree is mild), and there is a high requirement for solution accuracy, then the CLARABEL solver should be selected first. This solver has high numerical accuracy and stability when dealing with small- to medium-sized convex optimization problems.

[0094] If the problem is large and has high real-time requirements (i.e., remaining computation time) For smaller scheduling scenarios, the SCS solver is preferred. SCS uses a first-order split cone optimization algorithm, which is fast and suitable for large-scale, high-real-time scheduling scenarios.

[0095] If the constraint matrix has a condition number For larger matrices (i.e., highly ill-conditioned matrices) and with strict requirements for numerical stability, the ECOS solver should be preferred, and preconditioning should be enhanced to improve the matrix condition number and increase solution stability.

[0096] If the model contains nonlinear relaxation residuals If the value is large (i.e., the nonlinearity is strong), the CLARABEL solver is used in conjunction with a convexity iterative solution strategy to ensure the quality of the global optimal solution.

[0097] Next, the selected convex optimization solver is used to solve the cooperative scheduling model constructed in step S300. During the solution process, the objective function, constraints, and input data such as the predicted power obtained in step S200 are passed to the solver in a standard format. The solver iteratively calculates according to the interior-point method or the first-order cone optimization algorithm until it converges to the optimal solution.

[0098] If the solver successfully returns the optimal solution, the scheduling variables in the optimal solution (including the power grid interaction power, diesel engine output, energy storage charging and discharging power, etc. at each time moment) are used as the scheduling plan power output for use in step S500.

[0099] If an anomaly occurs during the solution process (such as solver non-convergence, numerical overflow, or reaching the iteration limit), a backup strategy is automatically activated. Specifically, the system presets a backup convex optimization solver, and after switching, uses the backup solver to resolve the same model. If the backup solver also returns an anomaly, the feasible scheduling plan from the previous time step is used for extrapolation correction to ensure stable output of scheduling instructions. Finally, the result obtained from a successful solution is used as the scheduling plan power. .

[0100] Through the aforementioned intelligent optimization mechanism of the dynamic solver, this embodiment can adaptively select the most suitable solution strategy based on the real-time operating status, ensuring solution accuracy while taking into account real-time performance and numerical stability, thus significantly improving the overall operating efficiency of the microgrid dispatching system.

[0101] like Figure 6 As shown, the power curve of the scheduling plan output by the microgrid 24-hour scheduling strategy model includes grid interaction power, diesel engine output and energy storage charging and discharging power.

[0102] Step S500: Calculate the scenario adaptive dynamic threshold based on the standard deviation of light intensity fluctuation, load factor and energy storage safety margin, obtain the actual executed power and compare it with the scheduling plan power to obtain the deviation. When the absolute value of the deviation exceeds the scenario adaptive dynamic threshold, correct the network parameters of the long short-term memory network in reverse according to the deviation, and re-execute the step of predicting the power generation of new energy units based on the corrected network parameters until the step of obtaining the scheduling plan power is obtained, and obtain the updated scheduling plan power.

[0103] Traditional "prediction-scheduling" closed-loop triggering mechanisms often employ fixed thresholds. This means that regardless of changing operating conditions, a correction process is triggered whenever the deviation between the actual power output and the scheduled power exceeds a certain constant value. However, microgrid operating conditions are complex and variable: in stable weather and with moderate loads, small deviations may stem from normal measurement noise, and frequent correction triggers would waste computational resources. Conversely, in extreme weather or when the energy storage state of charge (SOC) is nearing its limit, even small deviations can pose safety hazards, requiring a more sensitive response. Therefore, this embodiment proposes a scenario-adaptive dynamic threshold generation mechanism.

[0104] First, calculate the scene-adaptive dynamic threshold. This threshold is no longer a fixed value, but is dynamically adjusted based on current environmental fluctuations, load levels, and energy storage safety margins. The calculation formula is as follows:

[0105] in, For scene-adaptive dynamic thresholds, The base threshold is preset by the system administrator based on the overall control accuracy requirements of the microgrid. for The standard deviation of light intensity fluctuation over time is obtained by statistically analyzing the standard deviation of light intensity measurements over several past time windows. Larger fluctuations in light intensity indicate more unstable weather conditions, resulting in greater volatility in renewable energy output and requiring more sensitive deviation triggering conditions. for The load factor at any given moment is defined as the ratio of the current load power to the system's rated load. A higher load factor indicates that the system is operating near its economic point and has a lower tolerance for deviations. Conversely, a lower load factor provides the system with greater adjustment margin. for The constant-time energy storage safety margin is defined as the distance between the current State of Charge (SOC) and the safety boundary (such as the over-discharge or overcharge boundary). The smaller the margin, the more protection the energy storage device needs, and the threshold should be reduced accordingly to improve response sensitivity. This is an adaptive function that adjusts the threshold value according to fluctuations in sunlight, load levels, and energy storage status, used to scale the base threshold by considering these three factors. In this embodiment, the adaptive function... The specific form is as follows:

[0106] in, The standard deviation of light intensity fluctuation; , , These are weighting coefficients, obtained through training with historical data, reflecting the relative importance of each factor to scheduling decisions; This refers to the load factor.

[0107] Based on the dynamic threshold model described above, the response characteristics under different operating conditions are as follows: when the standard deviation of light intensity fluctuation... When the light intensity fluctuation is large (extreme weather), the threshold automatically tightens, the triggering frequency increases, and the system response becomes more sensitive; when the standard deviation of light intensity fluctuation is small (stable operating conditions), the threshold is appropriately relaxed to reduce unnecessary model updates and save computing resources; when the energy storage safety margin is large... When in a critical state, the threshold is minimized to maximize the safety of the energy storage device.

[0108] Before real-time monitoring of scheduling execution deviations, this invention first quantifies the output deviations of new energy scheduling to serve as the basis for subsequent reverse correction. The quantification expression is:

[0109] in, for The comprehensive quantitative value of the deviation in the dispatch output of new energy sources at any given time; for Predicted power generation error of new energy units at any time; It is in a state of energy storage charge; For load power; This represents the power grid operating state vector. The quantization function comprehensively considers prediction bias, energy storage status, load demand, and power grid operating conditions, providing more comprehensive bias characteristics for subsequent dynamic threshold determination and loss function construction.

[0110] Secondly, the system monitors scheduling execution deviations in real time and collects actual execution power in real time. and scheduling plan power Compare and calculate the absolute value of the deviation. Determine whether the deviation exceeds the scene adaptive dynamic threshold at the current moment. That is, to determine whether the following expression is true:

[0111] If the above conditions are met, the closed-loop correction process will be triggered.

[0112] like Figure 7 The diagram shown is a flowchart of the prediction-scheduling bidirectional coupling closed loop and the scheduling error reverse correction of this invention. The diagram clearly illustrates the complete self-healing closed loop from data acquisition, prediction, scheduling, execution, deviation monitoring to model parameter reverse correction and re-prediction.

[0113] Furthermore, once the closed-loop correction process is triggered, this invention constructs a backpropagation loss function based on the new energy dispatch output deviation and corrects the prediction model parameters in reverse. Specifically, a loss function driven by the new energy dispatch output deviation is constructed. :

[0114] in, For the power output deviation loss function of new energy dispatch, These are time-varying weighting coefficients; when the energy storage state of charge approaches the safety boundary, Automatically increase to enhance equipment protection and constraint satisfaction.

[0115] Subsequently, based on this loss function, the weight matrix of the Long Short-Term Memory network (including the forget gate weights) is corrected in reverse through an adaptive gradient descent mechanism. Input gate weights Output gate weights Memory unit weight and the attention weight coefficients of the feature attention mechanism The parameter update method is as follows:

[0116] in, Represented as The amount of network parameter updates at any given time, i.e., the magnitude of parameter adjustments. This indicates the network parameters to be updated. For learning rate, This represents the gradient of the loss function with respect to the network parameters. Through the aforementioned reverse correction, the prediction model can iterate and self-correct online based on the actual execution effect of the scheduling, thus suppressing prediction drift.

[0117] Finally, based on the corrected Long Short-Term Memory (LSTM) network parameters, steps S200 to S400 are re-executed: that is, the power generation of the new energy units is re-predicted to obtain the updated predicted power; then, the updated predicted power is input into the collaborative scheduling model to re-solve for the updated scheduling plan power. Through the above closed loop of "prediction-scheduling-execution-feedback-correction-re-prediction", the dynamic rolling update of the scheduling plan power is realized.

[0118] According to another aspect of the embodiments of this application, an electronic device is also provided, including a processor and a memory, wherein the processor is configured to implement the steps of the method when executing a computer program stored in the memory.

[0119] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0120] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0121] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0122] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0123] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A grid-energy storage coordinated dispatch method, characterized in that, The steps include the following: Step S100: Collect multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain standardized data; Step S200: Input the standardized data into a long short-term memory network with an embedded feature attention mechanism to predict the power generation of the new energy unit and obtain the predicted power. Step S300: With the goal of minimizing the total operating cost of the microgrid, set constraints, construct a three-layer bidirectional coupled collaborative scheduling model of day-ahead, intraday, and real-time, and input the predicted power into the collaborative scheduling model; Step S400: Collect real-time operating characteristics of the system to construct a multi-dimensional state vector, and adaptively select a convex optimization solver based on the multi-dimensional state vector to solve the cooperative scheduling model and obtain the scheduling plan power; Step S500: Calculate the scenario adaptive dynamic threshold based on the standard deviation of light intensity fluctuation, load factor and energy storage safety margin, obtain the actual executed power and compare it with the scheduling plan power to obtain the deviation. When the absolute value of the deviation exceeds the scenario adaptive dynamic threshold, correct the network parameters of the long short-term memory network in reverse according to the deviation, and re-execute the step of predicting the power generation of new energy units based on the corrected network parameters until the step of obtaining the scheduling plan power is obtained, and obtain the updated scheduling plan power.

2. The grid-energy storage coordinated dispatch method as described in claim 1, characterized in that, The collection of multi-source heterogeneous data includes: collecting historical power generation, load power, meteorological environment information, and energy storage data of new energy units as multi-source heterogeneous data. The meteorological environment information includes temperature, humidity, and light intensity, and the energy storage data includes battery state of charge, upper and lower limits of charging and discharging power, rated power of diesel engines, time-of-use electricity price of the power grid, and power grid operation status. The method for preprocessing the multi-source heterogeneous data includes: outlier removal, missing value imputation, and normalization of the multi-source heterogeneous data.

3. The grid-energy storage coordinated dispatch method as described in claim 1, characterized in that, The method of inputting the standardized data into a long short-term memory network with an embedded feature attention mechanism to predict the power generation of new energy units includes: Construct a Long Short-Term Memory (LSTM) network with an embedded feature attention mechanism; wherein the LSM network includes a forget gate, an input gate, an output gate, and a memory unit, and the feature attention mechanism is used to dynamically calculate attention weight coefficients for each input feature and to use the attention weight coefficients to perform weighted updates on the standardized data; The Long Short-Term Memory (LSTM) network with embedded feature attention mechanism is trained offline using preprocessed historical multi-source heterogeneous data. The mean squared error is used as the loss function, and the LSM network parameters are updated through backpropagation until the loss function converges, thus obtaining the trained LSM network. The data, which is collected in real time and standardized, is input into the trained Long Short-Term Memory network. After forward propagation calculation through the forget gate, input gate, output gate and memory unit, the predicted value of the power generation of the new energy unit is output as the predicted power.

4. The grid-energy storage coordinated dispatch method as described in claim 3, characterized in that, The method for constructing a three-layer bidirectional coupled collaborative scheduling model (day-ahead, intraday, and real-time) with constraints, aiming to minimize the total operating cost of microgrids, includes: Construct an objective function, which aims to minimize the total operating cost of the microgrid, and its expression is: in, for Power grid interaction at all times for Time-of-use electricity pricing at any given moment for The actual output of the diesel engine at any given time. This is the operating cost coefficient per unit power of the diesel engine; This represents the total number of scheduling periods; The constraints are defined, including energy balance constraints, energy storage battery state constraints, and device power constraints; the energy balance constraints are as follows: in, for Real-time new energy power forecast and They are respectively Real-time energy storage discharge power and charging power, for Real-time load power; The energy storage battery state constraints include the dynamic update equation for the state of charge and upper and lower bound constraints as follows: in, for Energy storage status at all times For battery capacity, For time intervals, Indicates the charging and discharging efficiency of the energy storage system; The power constraints of the equipment include upper and lower limits of charging and discharging power of the energy storage system and upper and lower limits of generating power of the diesel engine. in, This represents the lower limit of the actual output power of the diesel engine. This represents the upper limit of the actual output power of the diesel engine. This represents the lower limit of the discharge power of the energy storage system. This represents the upper limit of the discharge power of the energy storage system. The lower limit of the charging power for energy storage systems; The upper limit of the charging power of the energy storage system; A three-layer bidirectional coupled scheduling architecture is constructed, consisting of a day-ahead layer, an intraday layer, and a real-time layer. The day-ahead layer generates a baseline scheduling plan with a first period and a first step length, and transmits the baseline scheduling plan and prediction confidence interval to the intraday layer. The intraday layer receives deviation information uploaded by the real-time layer with a second period as the rolling window and a second step length, dynamically adjusts the scheduling trajectory, and sends a re-optimization command to the day-ahead layer when the accumulated deviation exceeds a preset threshold. The real-time layer outputs an instantaneous power correction command with a third period and a third step length, and feeds back the real-time deviation characteristics to the intraday layer.

5. The grid-energy storage coordinated dispatch method as described in claim 1, characterized in that, The method for constructing a multidimensional state vector based on the real-time operating characteristics of the acquisition system, and adaptively selecting a convex optimization solver based on the multidimensional state vector includes: The system collects real-time operational characteristics and constructs a multi-dimensional state vector, expressed as follows: in, It is a multidimensional state vector. for Number of decision variables at any given time for Number of time-bound constraints for The condition number of the constraint matrix at any given time. for Time-varying nonlinear relaxation residuals for The remaining time of the current scheduling cycle; The multidimensional state vector is input into a lightweight decision tree model trained offline, and the online performance of multiple convex optimization solvers is evaluated. The optimal solver is selected based on the evaluation results.

6. The grid-energy storage coordinated dispatch method as described in claim 5, characterized in that, The method for solving the cooperative scheduling model to obtain the scheduling plan power includes: applying the selected convex optimization solver to the cooperative scheduling model; if the solution returns the optimal solution, the optimal solution is output as the scheduling plan power; if the solution is abnormal, the method switches to a preset backup convex optimization solver to solve the cooperative scheduling model again, and the solution result of the backup convex optimization solver is used as the scheduling plan power.

7. The grid-energy storage coordinated dispatch method as described in claim 6, characterized in that, Based on the standard deviation of light intensity fluctuation, load factor, and energy storage safety margin, a scenario adaptive dynamic threshold is calculated. The actual executed power is obtained and compared with the scheduling plan power. When the absolute value of the deviation exceeds the scenario adaptive dynamic threshold, the following occurs: The scene-adaptive dynamic threshold is calculated using the following expression: in, For scene-adaptive dynamic thresholds, Based on the threshold, for Standard deviation of light intensity fluctuation over time for real-time load factor for Always have a safe and sufficient energy storage capacity. It is an adaptive function, which adjusts the threshold value according to the fluctuation of light intensity, load level and energy storage status; Real-time acquisition of actual execution power Obtain the scheduling plan power Calculate the absolute value of the deviation The deviation is compared with the scene adaptive dynamic threshold. If the absolute value of the deviation is greater than the scene adaptive dynamic threshold, the closed-loop correction process is triggered.

8. The grid-energy storage coordinated dispatch method as described in claim 7, characterized in that, The method for obtaining the updated scheduling plan power includes: Correcting the network parameters of the long short-term memory network in reverse based on the deviation, and re-executing the step of predicting the power generation of new energy units based on the corrected network parameters until the scheduling plan power is obtained. Based on the aforementioned deviation, a new energy dispatch output deviation loss function is constructed: in, The power output deviation loss function for new energy dispatching. These are time-varying weighting coefficients; when the energy storage state of charge approaches the safety boundary, Automatically increases in size; Using the new energy dispatch output deviation loss function, the weight matrix and feature attention weight coefficients of the long short-term memory network are corrected in reverse through an adaptive gradient descent mechanism to obtain the corrected long short-term memory network. Based on the modified Long Short-Term Memory network, the step of predicting the power generation of new energy units is re-executed to obtain the updated predicted power. Then, based on the updated predicted power, the steps of constructing the collaborative scheduling model and solving it are re-executed until the updated scheduling plan power is obtained.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the grid-energy storage coordinated scheduling methods of claims 1-8, and the processor is configured to execute the programs stored in the memory.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the grid-energy storage coordinated scheduling method according to any one of claims 1-8.

Images