Island energy storage multi-time scale soc optimization control method and system based on prediction uncertainty driving

Abstract

An island energy storage multi-time scale SOC optimization control method and system based on prediction uncertainty driving, comprising: collecting historical meteorological data and historical operation data of an island micro-grid system, and constructing a multi-dimensional time sequence feature set; based on the feature set, using a pre-trained uncertainty prediction model to quantify the prediction error and output the probability prediction interval and point prediction value of new energy output; based on the probability prediction interval, establishing a day-ahead scheduling optimization model, calculating the upper and lower limit constraints of the energy storage day-ahead reference output and standby capacity that meet the interval fluctuation extreme value condition; in the real-time stage, based on the deviation of the actual sampling power and the point prediction value, using a reinforcement learning algorithm to perform state optimization within the upper and lower limit constraints of the standby capacity, and dynamically outputting the energy storage system action instruction to complete the multi-time scale optimization control of the state of charge of the island micro-grid energy storage unit. The present application can guarantee the operation boundary safety of the island micro-grid system and improve its low-carbon economy.

Description

Technical Field

[0001] This invention relates to the field of microgrid energy management and automation control technology, specifically a method and system for multi-timescale SOC optimization control of islanded energy storage driven by predictive uncertainty. Background Technology

[0002] With the rapid development of renewable energy technologies, isolated microgrids are receiving increasing attention due to their ability to independently provide reliable power to remote areas and islands. In isolated operation mode, microgrids lose the voltage and frequency support of the main power grid, and the power balance within the system depends entirely on the coordinated operation of distributed power sources, local energy storage systems, and controllable loads.

[0003] Distributed renewable energy generation is characterized by strong randomness and non-stationary time-varying features. Especially in variable weather environments such as onshore wind gaps or offshore areas, wind speed and accompanying multidimensional meteorological variables exhibit nonlinear time-varying characteristics, which can easily lead to increased power fluctuations in islanded microgrids, making it difficult to balance source and load and control frequency. As a key device for smoothing power fluctuations and providing system inertia and backup, the reasonable control of the state of charge (SOC) of energy storage systems directly determines the operational safety and economy of islanded systems.

[0004] Existing microgrid dispatch schemes largely rely on deterministic point forecasting. However, traditional deterministic point forecasting methods cannot quantitatively characterize the probability distribution and tail risk of uncertainty errors. When faced with high-dimensional meteorological characteristics and complex microclimate disturbances, the generalization ability and characterization accuracy of point forecasts significantly decrease. If day-ahead dispatch plans are formulated solely based on point forecast trajectories, the reservation of energy storage reserve capacity may result in two situations: over-allocation will reduce the economic utilization efficiency of energy storage resources; under-allocation will fail to cope with the risk of power drop, potentially leading to wind curtailment, load shedding, and even affecting the stable operation of the microgrid. Furthermore, existing deep learning prediction models are insufficient in extracting the spatial correlation and temporal dependence features of multi-dimensional time-series meteorological data, and the model hyperparameters largely depend on manual experience settings, requiring improvement in the accuracy and reliability of the prediction interval.

[0005] At the multi-timescale SOC control level of energy storage, traditional mathematical programming-based energy management methods struggle to meet the demands of minute-level real-time control when dealing with high-dimensional uncertain nonlinear dynamic optimization problems. This is because decision variables expand exponentially with the control time domain. Conventional reinforcement learning methods based on unconstrained frameworks lack clear physical boundary constraints during the action exploration phase, easily leading to out-of-bounds action outputs. Frequent charging and discharging state switching can accelerate the capacity decay of energy storage batteries and adversely affect the transient stability of microgrids. Furthermore, existing methods generally lack mechanisms to directly couple the uncertainty of renewable energy output forecasting into energy storage SOC control, making it difficult to achieve multi-timescale collaborative optimization of SOC from day-ahead planning to real-time response.

[0006] To address the aforementioned issues, this invention proposes a method that integrates predictive uncertainty quantification with multi-timescale SOC optimization control. This method enables refined and robust SOC control of isolated microgrid energy storage units under complex meteorological conditions, thereby improving the system's renewable energy absorption capacity and operational stability. Summary of the Invention

[0007] The technical problem to be solved by this invention is: to address the strong uncertainty of new energy output caused by complex weather conditions in isolated microgrids, and the lack of multi-timescale coordination in the existing scheduling methods for energy storage state of charge (SOC) control, which makes it difficult to balance economic efficiency and physical safety boundaries. This invention provides an optimized control method and system for isolated energy storage with multi-timescale SOC based on prediction uncertainty.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] A method for multi-timescale SOC optimization control of isolated energy storage based on prediction uncertainty includes the following steps: S1, collecting historical meteorological data and historical operational data of the isolated microgrid system to construct a multi-dimensional time-series feature set; S2, based on the multi-dimensional time-series feature set, using a pre-trained uncertainty prediction model, quantifying the prediction error, and outputting the probability prediction interval and point prediction value of new energy output; S3, establishing a day-ahead scheduling optimization model based on the probability prediction interval, with the goal of minimizing the overall operating cost of the isolated microgrid, calculating the upper and lower limits of the day-ahead benchmark output and reserve capacity of energy storage that satisfy the extreme value condition of interval fluctuation; S4, in the real-time scheduling stage, based on the deviation between the actual sampled power and the point prediction value, using a reinforcement learning algorithm to perform strategy optimization within the upper and lower limits of the reserve capacity, outputting real-time action commands of the energy storage system to complete the multi-timescale optimization control of the state of charge (SOC) of the energy storage unit in the isolated microgrid.

[0010] Furthermore, in S1: the historical meteorological data includes at least one of wind speed, temperature, humidity, and air pressure; the historical operating data includes historical output data and load data of new energy units.

[0011] Furthermore, in S2: the uncertainty prediction model adopts a deep learning network based on quantile regression, specifically a QR-CNN-BiLSTM network; the QR-CNN-BiLSTM network extracts the spatial correlation of the multidimensional temporal feature set through a one-dimensional convolutional neural network, extracts time-dependent features through a bidirectional long short-term memory network, and is trained in a fully connected layer combined with a quantile loss function to output probability prediction intervals at multiple preset confidence levels;

[0012] During the offline training phase of the QR-CNN-BiLSTM model, the Capsule Monkey Search (CapSA) algorithm is used to optimize the network hyperparameters in order to minimize the fitting error.

[0013] Furthermore, in S3: the comprehensive operating cost of the day-ahead scheduling optimization model includes power generation cost, system source-load imbalance penalty cost, and tiered carbon trading cost; when determining the upper and lower limits of the day-ahead benchmark output and reserve capacity of the energy storage system, the day-ahead adjustable reserve capacity of the energy storage system is greater than or equal to the power fluctuation difference between the upper and lower limits of the probability prediction interval, so that the reserved energy storage capacity boundary covers the range of predicted uncertainty fluctuations.

[0014] Furthermore, in S4: the reinforcement learning algorithm adopts the Flexible Action-Evaluation (SAC) algorithm; in the real-time scheduling stage, the current output status of each unit in the islanded microgrid and the error value of the real-time power deviation from the predicted value are used as the environmental state input of the SAC algorithm, and the real-time energy storage charging and discharging power is used as the action space; the Flexible Action-Evaluation algorithm introduces action entropy, updates the network parameters according to the reward value fed back by the system, and outputs the real-time energy storage charging and discharging action command that maximizes the cumulative reward without breaking the upper and lower limits of the reserve capacity.

[0015] A multi-timescale SOC optimization control system for isolated energy storage driven by predictive uncertainty includes: a data acquisition module for collecting historical meteorological and operational data of the isolated microgrid system and constructing a multi-dimensional time-series feature set; an uncertainty prediction module for quantifying prediction errors and outputting the probability prediction interval and point prediction value of renewable energy output based on the multi-dimensional time-series feature set using a pre-trained uncertainty prediction model; a day-ahead scheduling module for establishing a day-ahead scheduling optimization model based on the probability prediction interval, aiming to minimize the overall operating cost of the isolated microgrid, and calculating the upper and lower limits of the day-ahead benchmark output and reserve capacity of energy storage that satisfy the extreme value conditions of interval fluctuations; and a real-time control module for optimizing strategies within the upper and lower limits of the reserve capacity using a reinforcement learning algorithm based on the deviation between the actual sampled power and the point prediction value during the real-time scheduling phase, and outputting real-time action commands for the energy storage system. Furthermore,

[0016] Furthermore, in the day-ahead scheduling module, the comprehensive operating cost of the day-ahead scheduling optimization model includes power generation cost, system source-load imbalance penalty cost, and tiered carbon trading cost; when the day-ahead scheduling module determines the upper and lower limits of the energy storage day-ahead benchmark output and reserve capacity, the day-ahead adjustable reserve capacity of the energy storage system is greater than or equal to the power fluctuation difference between the upper and lower bounds of the probability prediction interval.

[0017] Furthermore, in the real-time control module, the reinforcement learning algorithm adopts a flexible action-evaluation algorithm; the real-time control module uses the current output status of each unit in the islanded microgrid and the error value of the real-time power deviation from the predicted value as the environmental state input of the flexible action-evaluation algorithm, and uses the real-time energy storage charging and discharging power as the action space; the flexible action-evaluation algorithm introduces action entropy, updates network parameters based on the reward value fed back by the system, and outputs energy storage real-time charging and discharging action commands that maximize the cumulative discount reward without exceeding the upper and lower limits of the reserve capacity.

[0018] An electronic device is characterized by comprising a processor and a memory, wherein the processor is configured to execute a computer program stored in the memory to implement the islanded energy storage multi-timescale SOC optimization control method based on predictive uncertainty driven as described above.

[0019] A computer-readable storage medium is characterized in that the computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the multi-timescale SOC optimization control method for islanded energy storage based on predictive uncertainty driven as described above.

[0020] The beneficial effects of this invention are reflected in the following aspects:

[0021] The foundation for high-precision SOC optimization control: It abandons the traditional deterministic point prediction and normal distribution assumptions, and adopts a QR-CNN-BiLSTM network combined with multi-dimensional auxiliary meteorological variables to extract spatiotemporal features and directly outputs the probability prediction interval. This provides highly reliable uncertainty boundary information for multi-timescale optimization control of energy storage SOC. Compared with existing methods, it effectively improves the compactness and reliability of the prediction interval under the complex microclimate conditions of isolated islands.

[0022] Multi-timescale SOC collaborative control capability: By using the physical safety boundary reserved by the day-ahead scheduling and the SOC tracking reward mechanism of the real-time SAC algorithm, multi-timescale closed-loop control of SOC from hourly day-ahead planning to minute-level real-time adjustment is realized, which solves the problem of lack of day-ahead-real-time coordination in existing methods of SOC control.

[0023] Adaptive scheduling that balances low-carbon economy and SOC balance: A tiered carbon trading mechanism is introduced in the day-ahead phase to reduce overall carbon emissions; in the real-time phase, a SAC reinforcement learning algorithm with action entropy is used for multi-step state optimization, and an explicit SOC deviation penalty term is added to the reward function, so that the system can effectively improve economic efficiency while maintaining the energy storage SOC fluctuation within a reasonable range in a harsh isolated operating environment.

[0024] Excellent model generalization ability: This system is not only applicable to conventional onshore microgrids, but also maintains stable control performance under different datasets such as offshore wind conditions, and has broad prospects for engineering promotion. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 The flowchart is a main diagram of a multi-timescale SOC optimization control method for islanded energy storage based on prediction uncertainty driven by an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the topology and control architecture of an islanded microgrid system provided in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the uncertainty quantification prediction model structure based on QR-CNN-BiLSTM provided in an embodiment of the present invention.

[0029] Figure 4This invention provides a day-to-day to real-time multi-timescale linkage SOC optimization framework that takes into account the prediction interval. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] This invention proposes a multi-timescale SOC optimization control method for islanded energy storage driven by predictive uncertainty. The overall process is as follows: Figure 1 As shown. The following uses a typical islanded microgrid simulation model as an implementation vehicle to describe the implementation process of this invention in detail. Those skilled in the art should understand that the following parameters, operating conditions, and performance data are all obtained based on the simulation platform. In actual engineering applications, adjustments can be made according to specific scenarios to achieve similar technical effects. The method includes the following steps:

[0032] S1. Collect historical meteorological and operational data of the isolated microgrid system to construct a multi-dimensional time-series feature set. In addition to the core wind speed variable, extract multiple auxiliary meteorological variables (such as temperature) at historical clock times. Constructing multidimensional time-series input vectors .

[0033] S2. Construct a QR-CNN-BiLSTM prediction model based on quantile regression.

[0034] To quantify the uncertainty of new energy output and address the problem that traditional deterministic point prediction cannot characterize the probability distribution of prediction errors, this invention constructs a deep time series network based on quantile regression (QR).

[0035] This invention maps the spatial local correlations of multidimensional meteorological features through a one-dimensional convolutional neural network (CNN) layer, and then utilizes a bidirectional long short-term memory network (BiLSTM) to construct an analytical chain of long-sequence time-dependent features along bidirectional temporal evolution directions. The prediction model constructed in this invention does not output a single deterministic point prediction trajectory at the fully connected layer, but instead combines quantile regression theory to output multiple preset confidence levels. Quantile predictions below .

[0036] The optimization of network parameters and joint training of this prediction model are achieved by minimizing the quantile loss function, the mathematical derivation of which is as follows:

[0037] For a given quantile confidence level The actual sampled value of wind power is The corresponding quantile prediction value is Then the conditional quantile loss function Defined as:

[0038] .

[0039] To obtain a complete deterministic baseline prediction trajectory and a highly reliable probability interval, this invention selects a low quantile level. Median level and high quantile level Multi-task joint training is performed. Finally, the trained model outputs the predicted median point value. As the baseline input for SOC optimization control, it outputs a non-parametric probability prediction interval with a confidence level of 90%. This is used to quantify the uncertainty of prediction error. During the offline training phase of the model, the Cap-tailed Monkey Search (CapSA) algorithm is introduced to globally and adaptively optimize the network hyperparameters in order to minimize the overall performance index of quantile prediction error.

[0040] S3. Construct a day-ahead scheduling model driven by predictive uncertainty.

[0041] During the day-ahead dispatch period, the energy management system, based on the probability prediction intervals under the aforementioned multiple confidence levels, uses the comprehensive operating cost of the islanded microgrid. The minimum objective function is to formulate the day-ahead optimal output plan, unit combination scheme, and state of charge (SOC) reference trajectory for energy storage units. This reference trajectory will serve as the core reference for SOC equalization optimization during the real-time control phase. The mathematical model for the day-ahead robust optimization scheduling objective function is constructed as follows:

[0042] .

[0043] In the formula, This is the day-ahead scheduling cycle, with a value of 24, corresponding to a time step of 1 hour. The 24 scheduling moments unfolded; The operating, maintenance, and power generation costs of each distributed power source, controllable unit, and energy storage system within an independent microgrid; The tiered carbon trading costs are calculated in segments for the system. This refers to the penalty cost for wind curtailment or load shedding caused by source-load imbalance.

[0044] To effectively mitigate the risk of dynamic power imbalance in independent microgrids caused by stochastic prediction errors, the day-ahead dispatch model introduces a strongly coupled physical constraint between the prediction envelope boundary and the energy storage reserve capacity. Day-ahead dispatch stipulates that the adjustable reserve power margin reserved by the energy storage system must completely enclose the fluctuation limit of the prediction interval. Let the maximum permissible energy state of the energy storage system be... The minimum safe energy state is The results of recent optimization calculations The reference energy state of the energy storage system at any given time is Meanwhile, let the charging efficiency of the energy storage system be... The discharge efficiency is Then the day-ahead energy state and dynamic power reserve limit. and lower limit The following interval constraints must be strictly satisfied:

[0045] .

[0046] .

[0047] Using the above formula, day-ahead dispatch directly transforms the time-varying uncertainty range of non-parametric predictions into upper and lower limits of power regulation constraints for the energy storage system during the real-time control phase. This provides physical boundary conditions for the multi-timescale optimization control of energy storage SOC.

[0048] S4. Real-time adaptive control and SOC dynamic equilibrium under boundary constraints.

[0049] During the real-time control phase, independent microgrids face high-frequency, low-amplitude transient source-load bidirectional disturbances. This invention employs a Flexible Action-Evaluation (SAC) algorithm based on a maximum entropy reinforcement learning framework to perform adaptive multi-step dynamic optimization control within the day-ahead reserved power regulation margin, achieving real-time dynamic response and equilibrium recovery of the energy storage SOC. The key elements of its Markov decision process are defined as follows:

[0050] (1) State space The state space introduces a transient deviation between the real-time sampling power and the day-ahead prediction baseline to guide the agent in perceiving the evolutionary characteristics and degree of uncertainty fluctuations in real time. Its initial mathematical vector definition is:

[0051]

[0052] In the formula, This represents the actual output power of wind power at the real-time stage. This represents the current state of charge of the energy storage system. This represents the actual power demand of the local diverse loads. This is a scalar representation of the transient error between the actual sampled wind power and the day-ahead predicted median. It aims to eliminate the large numerical characteristics of wind power and load power, as well as the dimensionless state of charge. To address the interference of dimensional differences between the two systems on the stability of gradient updates in the policy network, this invention introduces a feature normalization function to perform real-time normalization mapping on the state space, completely restricting the elements of the state vector to within a certain range. Within the consistent interval.

[0053] (2) Action space The decision action is defined as the real-time charging and discharging control power of the energy storage system. To ensure the safety and lifespan of the energy storage medium, the operational output is strictly limited to the upper and lower limits of the reserve power calculated before the day, i.e., satisfying the following hard constraints:

[0054]

[0055] (3) Reward function The objective function is to minimize the real-time dynamic adjustment cost and the penalty for deviation from the state of charge. The reward function adopts a negative penalty form, and its expression is:

[0056]

[0057] In the formula, The real-time action adjustment cost introduced for transient frequency modulation and smoothing high-frequency power fluctuations is taken as in this embodiment. ; This is the reference trajectory for the energy storage state of charge established during the recent economic dispatch phase. and These are the real-time operation adjustment cost penalty coefficient and the state deviation penalty coefficient, respectively. Both of them are greater than zero, used to balance the economics of control and the flexibility margin of energy storage SOC state recovery.

[0058] (4) Maximum entropy optimization objective: The SAC algorithm introduces policy action entropy into the traditional objective function. This encourages agents to maintain sufficient exploratory capabilities in dynamic environments where they operate independently, avoiding the solution set from getting trapped in local optima. Policy Network The optimal decision control objective function Construct the expected sum of the cumulative discount reward and the policy action entropy:

[0059]

[0060] In the formula, The time-domain steps for real-time control correspond to minute-level real-time scheduling clock nodes; It is an adaptive temperature coefficient used to dynamically adjust the balance between strategy exploration and economic efficiency; This represents the distribution density of action state trajectories in the policy network under controlled conditions. Without exceeding physical boundary constraints, the agent dynamically outputs real-time charging and discharging action commands that maximize the cumulative discounted reward for energy storage, thereby achieving multi-timescale closed-loop optimization control of the energy storage SOC.

[0061] The process of constructing the detailed formula for comprehensive operating cost in S3 is as follows:

[0062] In the recent robust scheduling optimization model, the overall operating cost of independent microgrids was considered. System operation and maintenance and power generation costs Tiered carbon trading costs and the penalty cost for system source load imbalance It is composed of linear superposition. Its underlying mathematical sub-model is elaborated as follows:

[0063] (1) System operation and maintenance and power generation costs

[0064]

[0065] In the formula, the calculation formula for the operation and maintenance cost of distributed renewable energy is:

[0066]

[0067] in and These are the unit operation and maintenance cost coefficients for wind power systems and photovoltaic power arrays, respectively. The formula for calculating the comprehensive cost of fuel power generation and operation and maintenance for controllable units (such as micro gas turbines or diesel generators) is as follows:

[0068]

[0069] in Let be the output power of the controllable generator unit, and a, b, c be the quadratic function coefficients of the fuel consumption characteristics of the controllable generator unit. This represents the operation and maintenance loss factor per unit output power of the generating unit. The formula for calculating the lifespan loss cost of an energy storage system is as follows:

[0070]

[0071] in To calculate the cost factor for the physical lifetime degradation and capacity degradation of energy storage units per charge-discharge cycle, This refers to the charging and discharging power of the clock node in the energy storage system.

[0072] (2) Tiered carbon trading costs

[0073] The generational and credit of actual carbon emissions of microgrids during the day-ahead dispatch cycle Defined as the difference between the actual carbon emissions of controllable units and the carbon allowances allocated free of charge by the system:

[0074]

[0075] In the formula, Carbon emission intensity coefficient per unit of electricity generated; This refers to the scalar amount of free basic carbon emission allowances allocated to independent microgrids. To eliminate mathematical ambiguity caused by ellipses in the segmentation, this invention constructs a general formula for recursive cumulative tiered carbon trading costs. Let the actual carbon emission allowance fall within the [missing information - likely a specific value or category]. Within each stepped interval, the constraint conditions are satisfied.

[0076]

[0077] Let the length of the tier interval be l (unit: tons), and the tier increase coefficient be α (dimensionless, α > 0), then the carbon trading cost of the nth tier is:

[0078]

[0079] In the formula, It serves as the benchmark carbon trading price for the regional electricity carbon trading market; The threshold for the length of the carbon emission ladder interval; This represents the tiered carbon price penalty increase rate after exceeding the baseline allowance range. Note: When the actual carbon emissions do not exceed the first tier (i.e., When ), the preceding cumulative term Automatically zeroed out, this general term formula is completely equivalent to the benchmark linear carbon cost formula.

[0080]

[0081] (3) System source-load imbalance penalty cost

[0082]

[0083] In the formula, For the system in The wind power curtailment caused by insufficient power absorption margin; This refers to the load shedding power caused by the depletion of the system's reserve capacity. and These are the penalty weighting coefficients for the corresponding unit wind curtailment and unit load shedding, respectively. Their values ​​are usually set relatively large to reflect the robust safety orientation of the dispatching scheme.

[0084] This embodiment uses a specific isolated microgrid as an example for illustration, and its topology and control architecture are as follows: Figure 2 As shown, the model includes wind turbine generators (WT), photovoltaic arrays (PV), controllable diesel generator sets (MT / CG), and lithium iron phosphate energy storage systems (BESS). Each unit is connected to the islanded microgrid AC main bus (ACBus) via a bidirectional converter or active power control unit.

[0085] S1. Collect historical meteorological and operational data of the isolated microgrid system to construct a multi-dimensional time-series feature set. Specifically, the Information Control System (EMS) collects data on the operational status of multi-source meteorological monitoring stations and each unit through industrial communication protocols.

[0086] This embodiment uses typical isolated microgrid operation data, spanning 90 days with a sampling interval of 15 minutes. The collected data undergoes cleaning and preprocessing: [The following is a list of methods used in the original text, which is not translated here.] Outliers were removed in principle, and missing values ​​were imputed using linear interpolation. The preprocessed data was divided into training (70%), validation (15%), and test (15%) sets in chronological order, maintaining temporal continuity during the division. Then, the minimum value for each feature was calculated based solely on the training set. and maximum value The training set is then normalized using Min-Max normalization, mapping it to the [0,1] interval. The normalization formula is as follows:

[0087]

[0088] Subsequently, using the same and The parameters are normalized in the same way for both the validation set and the test set.

[0089] S2. Uncertainty prediction based on hyperparameter optimization.

[0090] The prediction network employs a cascaded structure of a one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (BiLSTM), as illustrated in the diagram below. Figure 3As shown in the diagram, the CNN layer is used to extract the spatial correlation of meteorological features, and the BiLSTM layer is used to extract the temporal dependency. The network combines quantile regression in the fully connected layer, simultaneously outputting wind speed predictions at the low quantile (τ=0.05), median (τ=0.50), and high quantile (τ=0.95), thus forming a probability prediction interval with a 90% confidence level.

[0091] During the offline training phase, the Capsule Monkey Search (CapSA) algorithm is introduced to optimize network hyperparameters (such as kernel size, number of hidden neurons, and learning rate). With the goal of minimizing the fitness function (weighted sum of quantile losses), the predicted interval after training can cover the range of actual output fluctuations, providing boundary information for subsequent scheduling.

[0092] S3. SOC optimization driven by day-ahead uncertainty.

[0093] like Figure 4 The day-ahead to real-time multi-timescale linkage SOC optimization framework shown is divided into two stages: day-ahead scheduling and real-time control, taking into account the prediction interval. In the day-ahead stage, day-ahead scheduling uses a 1-hour time step and a rolling optimization cycle of 24 hours. The predicted values ​​of τ=0.05 and τ=0.95 quantiles output from step S2 are used as interval boundaries and substituted into the day-ahead scheduling optimization model. The comprehensive operating cost includes power generation cost, tiered carbon trading cost, and source-load imbalance penalty cost. Cost parameters can be set based on islanded engineering experience. Each cost parameter can be set with example values ​​based on typical islanded engineering experience (e.g., wind power operation and maintenance cost coefficient 50 yuan / MWh, diesel generator fuel cost coefficient a=0.01 yuan / kW²·h, energy storage life loss cost coefficient 100 yuan / MWh, benchmark carbon price 50 yuan / ton, wind curtailment penalty 800 yuan / MWh, load shedding penalty 2000 yuan / MWh, etc.).

[0094] The day-ahead optimization model is solved using a mixed-integer linear programming solver (such as Gurobi or CPLEX) to calculate the day-ahead baseline power output plan and SOC baseline reference trajectory of the energy storage system over 24 hours, while strictly locking the upper and lower limits of the energy storage reserve capacity for each time period. The reserve capacity constraint is set to at least cover the power fluctuation difference between the upper and lower boundaries of the prediction interval, thus reserving sufficient physical adjustment margin for real-time control. In the simulation, the SOC baseline trajectory is usually limited to a reasonable range (e.g., 0.2–0.8) to avoid overcharging or over-discharging.

[0095] S4, Real-time Adaptive Control and SOC Dynamic Equilibrium.

[0096] Real-time scheduling is refreshed every 5 minutes. A Flexible Action-Evaluation (SAC) reinforcement learning algorithm is employed to optimize the state within the day-ahead reserved reserve capacity boundary. The state space includes: actual wind power, current energy storage SOC, load power, and the deviation between the actual power and the day-ahead predicted value. The action space is the energy storage charging and discharging power, and its output is limited to the upper and lower limits of the reserve capacity calculated day-ahead. The reward function includes a real-time adjustment cost term and a SOC deviation penalty term, the latter ensuring that the energy storage SOC tracks the day-ahead generated baseline reference trajectory.

[0097] The SAC algorithm's network structure (Actor and Critic networks) employs fully connected hidden layers (e.g., 3 layers, 128 neurons per layer) with ReLU activation. Hyperparameters such as discount factor, learning rate, experience pool size, and temperature coefficient use configurations common in simulation environments (e.g., γ=0.99, learning rate 0.0003, experience pool size 1e6). Training is performed offline in a simulation environment, iteratively updating network parameters using historical data until the reward value converges.

[0098] In the exemplary real-time control process, the SAC agent outputs charge and discharge commands within the reserve capacity boundary. In this embodiment, the energy storage SOC can track the day-ahead reference trajectory, and the action commands do not exceed the safety boundary. The system frequency and voltage can be maintained within allowable ranges under typical disturbance conditions (e.g., frequency fluctuation within ±0.2Hz, voltage deviation within ±2%), verifying the effectiveness of the method of the present invention.

[0099] To verify the adaptability of this invention in scenarios with increased prediction uncertainty, a "sudden drop in wind speed" is used as an example: Imagine a moment when the wind speed drops significantly, causing a decrease in actual wind power output. The probability prediction range output by the prediction module widens, and the day-ahead dispatch reserves a larger energy storage reserve capacity during this period. During the real-time control phase, the SAC agent senses the deviation between the actual power and the predicted value, and increases the energy storage discharge power within the reserve capacity constraint to compensate for the power shortfall. In this embodiment, the system power balance is not disrupted, the frequency and voltage remain stable, and the SOC recovers to near the baseline trajectory after the disturbance.

[0100] The method described in this invention can be deployed on the edge computing server or industrial control computer of an isolated microgrid energy management system (EMS). Typical hardware configurations include: a multi-core CPU (for data acquisition and communication), an optional GPU (for neural network inference acceleration, not required), ≥16GB of memory, a solid-state drive, and dual gigabit Ethernet ports. The software operating environment can be based on Linux or Windows operating systems, using the Python programming language and mainstream deep learning frameworks (TensorFlow or PyTorch). The method described in this invention achieves data interaction with the underlying energy storage converter (PCS) and the microgrid central controller through standard industrial communication protocols (such as Modbus TCP / IP, IEC 61850). The communication cycle can be set to 200ms (the value used in this embodiment) to meet real-time control requirements.

[0101] This embodiment demonstrates that the method proposed in this invention can achieve the following effects:

[0102] By using probability prediction intervals to quantify the uncertainty of new energy output, a robust boundary is provided for energy storage SOC scheduling;

[0103] The scheduling and real-time control are linked to achieve collaborative optimization of SOC across multiple time scales;

[0104] Reinforcement learning algorithms make autonomous decisions within physical security boundaries, balancing economy and SOC equilibrium.

[0105] Under typical islanding disturbance conditions, the system frequency and voltage remained stable, and the energy storage SOC did not exceed the limit or deviate significantly.

[0106] Based on the content described in this specification, combined with conventional simulation tools and engineering experience, those skilled in the art can implement the method proposed in this invention and expect to obtain similar technical effects.

[0107] This invention also provides a multi-timescale SOC optimization control system for isolated energy storage driven by predictive uncertainty, comprising:

[0108] The data acquisition module is used to collect historical meteorological data and historical operational data of isolated microgrid systems to construct a multi-dimensional time-series feature set;

[0109] The uncertainty prediction module is used to quantify the prediction error and output the probability prediction interval and point prediction value of new energy output based on the multi-dimensional time series feature set and a pre-trained uncertainty prediction model.

[0110] The day-ahead scheduling module is used to establish a day-ahead scheduling optimization model based on the probability prediction interval, with the goal of minimizing the overall operating cost of the islanded microgrid, and to calculate the upper and lower limits of the energy storage day-ahead benchmark output and reserve capacity that meet the extreme value conditions of the interval fluctuation.

[0111] Real-time control module: During the real-time scheduling phase, based on the deviation between the actual sampled power and the predicted value at the point, a reinforcement learning algorithm is used to optimize the strategy within the upper and lower limits of the reserve capacity, and output the real-time action command of the energy storage system.

[0112] This invention also provides an electronic device, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the islanded energy storage multi-timescale SOC optimization control method based on predictive uncertainty driven as described above.

[0113] This invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the above-described method for multi-timescale SOC optimization control of islanded energy storage based on predictive uncertainty.

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-timescale SOC optimization control method for islanded energy storage driven by prediction uncertainty, characterized in that, Includes the following steps: S1. Collect historical meteorological data and historical operational data of the isolated microgrid system to construct a multi-dimensional time-series feature set; S2. Based on the multi-dimensional time series feature set, a pre-trained uncertainty prediction model is used to quantify the prediction error and output the probability prediction interval and point prediction value of new energy output. S3. Based on the probability prediction interval, establish a day-ahead scheduling optimization model. With the goal of minimizing the overall operating cost of the islanded microgrid, calculate the upper and lower limits of the day-ahead benchmark output and reserve capacity of energy storage that meet the extreme value conditions of the interval fluctuation. S4. During the real-time scheduling phase, based on the deviation between the actual sampled power and the predicted value at the point, a reinforcement learning algorithm is used to optimize the strategy within the upper and lower limits of the reserve capacity, and output the real-time action command of the energy storage system to complete the multi-time-scale optimization control of the state of charge (SOC) of the energy storage unit in the islanded microgrid.

2. The method for multi-timescale SOC optimization control of isolated energy storage based on prediction uncertainty driven by claim 1, characterized in that, In S1: The historical meteorological data includes at least one of wind speed, temperature, humidity, and air pressure; the historical operating data includes historical output data and load data of new energy units.

3. The method for multi-timescale SOC optimization control of isolated energy storage based on prediction uncertainty driven according to claim 1, characterized in that, In S2: The uncertainty prediction model employs a deep learning network based on quantile regression, specifically a QR-CNN-BiLSTM network. The QR-CNN-BiLSTM network extracts the spatial correlation of the multi-dimensional temporal feature set through a one-dimensional convolutional neural network, extracts time-dependent features through a bidirectional long short-term memory network, and is trained in a fully connected layer with a quantile loss function to output probability prediction intervals at multiple preset confidence levels. During the offline training phase of the QR-CNN-BiLSTM model, the Capsule Monkey Search (CapSA) algorithm is used to optimize the network hyperparameters in order to minimize the fitting error.

4. The method for multi-timescale SOC optimization control of isolated energy storage based on prediction uncertainty driven according to claim 1, characterized in that, In S3: The comprehensive operating cost of the day-ahead scheduling optimization model includes power generation cost, system source-load imbalance penalty cost, and tiered carbon trading cost; When determining the upper and lower limits of the day-ahead baseline output and reserve capacity of the energy storage system, the day-ahead adjustable reserve capacity of the energy storage system is greater than or equal to the power fluctuation difference between the upper and lower limits of the probability prediction interval, so that the reserved energy storage capacity boundary covers the range of predicted uncertainty fluctuations.

5. The method for multi-timescale SOC optimization control of isolated energy storage based on prediction uncertainty driven by claim 1, characterized in that, In S4: The reinforcement learning algorithm employs the Flexible Action-Evaluation (SAC) algorithm; During the real-time scheduling phase, the current output status of each unit in the islanded microgrid and the error value of the real-time power deviation from the predicted value are used together as the environmental state input of the SAC algorithm, and the real-time energy storage charging and discharging power is used as the action space. The flexible action-evaluation algorithm introduces action entropy and updates network parameters based on the reward value fed back by the system. Without exceeding the upper and lower limits of the reserve capacity, it outputs real-time charging and discharging action commands that maximize the cumulative discount reward.

6. A multi-timescale SOC optimization control system for islanded energy storage driven by predictive uncertainty, characterized in that, include: The data acquisition module is used to collect historical meteorological data and historical operational data of isolated microgrid systems to construct a multi-dimensional time-series feature set; The uncertainty prediction module is used to quantify the prediction error and output the probability prediction interval and point prediction value of new energy output based on the multi-dimensional time series feature set and a pre-trained uncertainty prediction model. The day-ahead scheduling module is used to establish a day-ahead scheduling optimization model based on the probability prediction interval, with the goal of minimizing the overall operating cost of the islanded microgrid, and to calculate the upper and lower limits of the energy storage day-ahead benchmark output and reserve capacity that meet the extreme value conditions of the interval fluctuation. Real-time control module: During the real-time scheduling phase, based on the deviation between the actual sampled power and the predicted value at the point, a reinforcement learning algorithm is used to optimize the strategy within the upper and lower limits of the reserve capacity, and output the real-time action command of the energy storage system.

7. The islanded energy storage multi-timescale SOC optimization control system based on prediction uncertainty driven according to claim 6, characterized in that: In the day-ahead scheduling module, the comprehensive operating cost of the day-ahead scheduling optimization model includes power generation cost, system source-load imbalance penalty cost, and tiered carbon trading cost; When the day-ahead scheduling module determines the upper and lower limits of the day-ahead reference output and reserve capacity of the energy storage system, the day-ahead adjustable reserve capacity of the energy storage system is greater than or equal to the power fluctuation difference between the upper and lower bounds of the probability prediction interval.

8. The islanded energy storage multi-timescale SOC optimization control system based on prediction uncertainty driven according to claim 6, characterized in that: In the real-time control module, the reinforcement learning algorithm adopts a flexible action-evaluation algorithm; The real-time control module uses the current output status of each unit in the islanded microgrid and the error value of the real-time power deviation from the predicted value as the environmental state input of the flexible action-evaluation algorithm, and uses the real-time energy storage charging and discharging power as the action space. The flexible action-evaluation algorithm introduces action entropy and updates network parameters based on the reward value fed back by the system. Without exceeding the upper and lower limits of the reserve capacity, it outputs real-time charging and discharging action commands that maximize the cumulative discount reward.

9. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the islanded energy storage multi-timescale SOC optimization control method based on prediction uncertainty driven as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the islanded energy storage multi-timescale SOC optimization control method based on predictive uncertainty driven as described in any one of claims 1 to 5.

Images