Method, system and device for evaluating multi-scenario regulation capability of energy storage power station and medium

Abstract

A kind of energy storage power station multi-scenario regulation capability evaluation method, system, equipment and medium, comprising: obtaining the actual operation data of energy storage power station facing peak regulation, frequency modulation and emergency power support scene;Actual operation data is input into the pre-constructed energy storage multi-scenario regulation capability comprehensive evaluation model, and the energy storage multi-scenario regulation capability comprehensive evaluation model is converted into Markov decision process;The Markov decision process is solved using the DDQN algorithm improved based on attention mechanism, and the comprehensive score of the energy storage multi-scenario regulation capability is obtained;The application realizes the collaborative evaluation of multiple scenes such as peak regulation, frequency modulation and emergency power support by constructing a multi-dimensional evaluation model;By converting the evaluation model into Markov decision process, the time sequence dependency in energy storage operation can be effectively handled, the multi-scenario evaluation accuracy is improved, dynamic and online evaluation is realized, and the dependence on historical data is avoided.

Description

Technical Field

[0001] This application relates to the field of regulation capacity assessment technology, specifically to a method, system, equipment and medium for assessing the regulation capacity of energy storage power stations in multiple scenarios. Background Technology

[0002] Driven by the current new power system and the "dual carbon" target, lithium-ion battery energy storage has become a key means to improve grid flexibility, and its inclusion in the 2024 government work report marks a new stage of large-scale application. However, although the installed capacity of new energy storage reached 44.44 million kilowatts in the first half of 2024, the equivalent utilization hours were only 390 hours, exposing prominent problems such as incomplete state perception and unclear boundaries of multi-scenario regulation capabilities, which seriously restricts its actual effectiveness. Therefore, establishing a precise evaluation method for peak shaving, frequency regulation, and emergency power support is of great theoretical and engineering significance for improving the dispatchability of energy storage and the reliability of the power grid.

[0003] Currently, domestic and international research has made some progress in energy storage assessment index systems, assessment model construction, and intelligent assessment methods, but certain limitations exist: existing literature only focuses on peak-shaving scenarios to construct strategies and assessment models, without addressing functions such as frequency regulation; existing literature assesses power plant performance from a comprehensive operational perspective, but does not deeply analyze its regulation capabilities in specific scenarios such as peak-shaving and frequency regulation; while existing literature has verified the synergistic advantages of hybrid energy storage in peak-shaving and frequency regulation, it has not systematically assessed the capability correlation and comprehensive benefits across multiple scenarios. Existing literature has constructed real-time performance assessment systems and optimized allocation strategies, but has not extended to multi-scenario collaborative assessments such as peak-shaving, nor has it resolved the multi-objective conflict problem under high-dimensional uncertainty. In summary, the limitations of existing research are concentrated in the following three points: First, domestic and international assessments of energy storage regulation capabilities are mostly focused on single scenarios, such as frequency regulation or peak-shaving, lacking a comprehensive assessment system for multi-scenario collaboration; second, existing assessment methods are mostly based on historical data statistical analysis, failing to incorporate real-time status monitoring data, resulting in lagging assessment results and limited accuracy; third, many assessment methods rely on offline experimental data or static models, making it difficult to handle multi-objective conflicts and high-dimensional uncertainties in actual operation.

[0004] Scheduling models incorporating energy storage optimization typically manifest as complex nonlinear programming problems, requiring the handling of multiple structural variables and nonlinear constraints, making modeling and solving them challenging. Furthermore, in assessing the multi-scenario regulation capabilities of energy storage, the problem essentially boils down to real-time state awareness and dynamic decision-making under complex operating environments, placing higher demands on the adaptability and convergence of the methods. To address this issue, deep Q-networks have gradually become a research hotspot due to their powerful state abstraction and policy generalization capabilities. Existing literature has achieved maximum revenue and minimum deviation by constructing an interactive response framework between energy storage virtual power plants and distributed resources, and formulating peak-shaving prices based on deep Q-networks (DQN). However, it has not fully considered the impact of battery life, and the long-term economic efficiency needs to be improved. Existing literature has designed energy storage management systems based on DQN, introducing charge-discharge cycle limits to extend battery life. It has been effective in smoothing grid fluctuations and providing power support, but it does not make sufficient use of time-series state information and lacks the ability to dynamically perceive and focus on key operating states. Existing literature has enhanced the ability to extract historical states and key operating information by introducing an attention mechanism based on the DQN framework, effectively improving the utilization rate and economy of energy storage. However, its basic framework is still based on DQN, and there is a problem of overestimation of Q value, which affects the stability of the strategy and long-term performance. Existing literature has proposed a dual-competition deep Q-network method based on priority playback for the grid edge control problem of community energy storage systems. By introducing a competition architecture and a priority experience playback mechanism, it has significantly improved the convergence speed and control accuracy. However, this method still relies on traditional neural network structures for action space discretization and state representation, and has limited ability to dynamically perceive temporal dependencies and key state features.

[0005] In summary, existing research on the multi-scenario coordinated control evaluation of battery energy storage systems (BESS) still suffers from problems such as simplistic strategies, insufficient parameter consideration, and weak utilization of real-time information. Summary of the Invention

[0006] To address the shortcomings of existing technologies in the multi-scenario coordinated control evaluation of BESS (Body Safety System), such as simplistic strategies, insufficient parameter consideration, and weak utilization of real-time information, this application proposes a method for evaluating the multi-scenario regulation capability of energy storage power stations, characterized by including: Acquire actual operational data of energy storage power stations in scenarios involving peak shaving, frequency regulation, and emergency power support; The actual operating data is input into a pre-constructed comprehensive evaluation model of energy storage multi-scenario regulation capability, and the comprehensive evaluation model of energy storage multi-scenario regulation capability is transformed into a Markov decision process. The Markov decision process is solved using a dual deep Q-network algorithm based on an improved attention mechanism to obtain a comprehensive score of the energy storage multi-scenario regulation capability. The comprehensive evaluation model for energy storage multi-scenario regulation capability is constructed by combining constraints with the optimization objectives of minimizing comprehensive cost and maximizing multi-scenario performance response.

[0007] Preferably, the construction process of the comprehensive evaluation model for the multi-scenario regulation capability of energy storage includes: Based on the actual operation data of energy storage power stations in various scenarios during historical periods, and combined with various evaluation indicators, the weights of each evaluation indicator are calculated using the analytic hierarchy process and the entropy weight method to obtain the comprehensive weight of each evaluation indicator. Based on each evaluation indicator and its comprehensive weight, the optimization objectives are to minimize the overall cost and maximize the performance response in multiple scenarios. The constraints are based on charging and discharging power constraints and state of charge constraints. A comprehensive evaluation model is constructed based on the optimization objective and the constraints. The evaluation metrics include: scenario power deviation, response rate, operating cost, degradation cost, state of charge (SOC), and charge / discharge power.

[0008] Preferably, the calculation formula for the optimization objective is as follows:

[0009] In the formula, The objective function is... For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. The overall weighting of safety indicators; For response rate;

[0010] In the formula, For deviation indicators in multiple scenarios; The comprehensive weighting of the deviation in peak shaving scenarios; For power deviation in peak shaving scenarios; The overall weight for frequency modulation scene deviation; For power deviation in frequency modulation scenarios; The overall weighting for deviations in emergency power support scenarios; The overall weight for deviations in emergency power support scenarios.

[0011] Preferably, the step of transforming the comprehensive evaluation model of energy storage multi-scenario regulation capability into a Markov decision process includes: Based on the actual operating data and the constraints of the comprehensive evaluation model for the multi-scenario regulation capability of energy storage, the state vector is selected to determine the state space; The action space is determined based on the energy storage charging and discharging power. The state transition probability is determined based on the probability of transitioning to the next state after taking an action in the current state; The reward function is determined by the reciprocal of the objective function of the comprehensive evaluation model of energy storage multi-scenario regulation capability. The reward discount factor is determined based on the time scale of peak shaving, frequency regulation, and emergency power support scenarios. The quintuple of a Markov decision process is determined based on the state space, action space, state transition probability, reward function, and reward discount factor.

[0012] Preferably, the step of solving the Markov decision process using a dual deep Q-network algorithm based on an attention mechanism to obtain a comprehensive score of the energy storage multi-scenario regulation capability includes: The Markov decision process is solved using a dual deep Q-network algorithm based on an improved attention mechanism to obtain the optimal charge-discharge power sequence; The values ​​of each evaluation index are obtained by calculating based on the optimal charge and discharge power sequence and various evaluation indicators. The comprehensive score of energy storage's multi-scenario adjustment capability is obtained by substituting the values ​​of each evaluation indicator into the scoring formula.

[0013] Preferably, the step of using an improved dual deep Q-network algorithm based on an attention mechanism to solve the Markov decision process and obtain the optimal charge-discharge power sequence includes: The state space of the Markov decision process is weighted by an attention mechanism to obtain the attention-weighted state. The attention-weighted state is input into the main network of the improved dual deep Q-network algorithm to select the current optimal action; The target network, based on the improved dual-depth Q-network algorithm, evaluates the current optimal action to obtain the Q value and updates the main network parameters until the algorithm converges, thus obtaining the optimal strategy and the optimal charge / discharge power sequence.

[0014] Preferably, the attention-weighted state obtained by performing attention-based weighting on the state space of the Markov decision process includes: The state space of the Markov decision process is normalized to obtain a normalized state vector. The global context vector is calculated based on the normalized feature vector; The similarity score is calculated based on the global context vector. Normalized attention weights are calculated based on similarity scores to obtain normalized attention weights. An attention weight matrix is ​​generated based on the normalized attention weights; The attention-weighted state is obtained by weighting the attention weight matrix and the normalized state vector.

[0015] Preferably, the calculation formula for the scoring formula is as follows:

[0016]

[0017] In the formula, The negative impacts of standardizing energy storage power stations; For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. for The number of times a scenario response is effective at any given moment; This is the final positive composite score.

[0018] Based on the same concept, this application also proposes a multi-scenario regulation capability assessment system for energy storage power stations, including: The data acquisition module is used to acquire actual operating data of energy storage power stations in scenarios such as peak shaving, frequency regulation, and emergency power support. The model conversion module is used to input the actual operating data into a pre-built comprehensive evaluation model of the energy storage multi-scenario regulation capability, and to convert the comprehensive evaluation model of the energy storage multi-scenario regulation capability into a Markov decision process. The algorithm solving module is used to solve the Markov decision process using a dual deep Q-network algorithm based on an improved attention mechanism, and obtain a comprehensive score of the energy storage multi-scenario regulation capability. The comprehensive evaluation model for energy storage multi-scenario regulation capability is constructed by combining constraints with the optimization objectives of minimizing comprehensive cost and maximizing multi-scenario performance response.

[0019] Preferably, it also includes a model building module, which is specifically used for: Based on the actual operation data of energy storage power stations in various scenarios during historical periods, and combined with various evaluation indicators, the weights of each evaluation indicator are calculated using the analytic hierarchy process and the entropy weight method to obtain the comprehensive weight of each evaluation indicator. Based on each evaluation indicator and its comprehensive weight, the optimization objectives are to minimize the overall cost and maximize the performance response in multiple scenarios. The constraints are based on charging and discharging power constraints and state of charge constraints. A comprehensive evaluation model is constructed based on the optimization objective and the constraints. The evaluation metrics include: scenario power deviation, response rate, operating cost, degradation cost, state of charge (SOC), and charge / discharge power.

[0020] Preferably, the optimization objective in the model building module is calculated as follows:

[0021] In the formula, The objective function is... For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. The overall weighting of safety indicators; For response rate;

[0022] In the formula, For deviation indicators in multiple scenarios; The comprehensive weighting of the deviation in peak shaving scenarios; For power deviation in peak shaving scenarios; The overall weight for frequency modulation scene deviation; For power deviation in frequency modulation scenarios; The overall weighting for deviations in emergency power support scenarios; The overall weight for deviations in emergency power support scenarios.

[0023] Preferably, the model conversion module is specifically used for: Based on the actual operating data and the constraints of the comprehensive evaluation model for the multi-scenario regulation capability of energy storage, the state vector is selected to determine the state space; The action space is determined based on the energy storage charging and discharging power. The state transition probability is determined based on the probability of transitioning to the next state after taking an action in the current state; The reward function is determined by the reciprocal of the objective function of the comprehensive evaluation model of energy storage multi-scenario regulation capability. The reward discount factor is determined based on the time scale of peak shaving, frequency regulation, and emergency power support scenarios. The quintuple of a Markov decision process is determined based on the state space, action space, state transition probability, reward function, and reward discount factor.

[0024] Preferably, the algorithm solving module includes: The Markov decision process solving submodule is used to solve the Markov decision process using a dual deep Q-network algorithm based on an attention mechanism to obtain the optimal charge and discharge power sequence. The index calculation submodule is used to calculate the value of each evaluation index based on the optimal charge and discharge power sequence and each evaluation index. The scoring calculation submodule is used to calculate the comprehensive score of the energy storage's multi-scenario adjustment capability by substituting the values ​​of each evaluation index into the scoring formula.

[0025] Preferably, the Markov decision process solver submodule includes: The attention weighting unit is used to perform attention mechanism weighting processing on the state space of the Markov decision process to obtain the attention-weighted state. The optimal action selection unit is used to input the attention-weighted state into the main network of the improved dual deep Q network algorithm to select the current optimal action; The optimal charge / discharge power sequence calculation unit is used to evaluate the current optimal action based on the target network of the improved dual-depth Q-network algorithm to obtain the Q value, and update the main network parameters until the algorithm converges, thereby obtaining the optimal strategy and the optimal charge / discharge power sequence.

[0026] Preferably, the attention weighting unit is specifically used for: The state space of the Markov decision process is normalized to obtain a normalized state vector. The global context vector is calculated based on the normalized feature vector; The similarity score is calculated based on the global context vector. Normalized attention weights are calculated based on similarity scores to obtain normalized attention weights. An attention weight matrix is ​​generated based on the normalized attention weights; The attention-weighted state is obtained by weighting the attention weight matrix and the normalized state vector.

[0027] Preferably, the calculation formula in the scoring calculation submodule is as follows:

[0028]

[0029] In the formula, The negative impacts of standardizing energy storage power stations; For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. for The number of times a scenario response is effective at any given moment; This is the final positive composite score.

[0030] Furthermore, this application also proposes an electronic device comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a method for evaluating the multi-scenario regulation capability of an energy storage power station as described above is implemented.

[0031] Furthermore, this application also proposes a readable storage medium on which an executable program is stored, which, when executed, implements the multi-scenario regulation capability assessment method for energy storage power stations as described above.

[0032] Compared with the prior art, the beneficial effects of this application are as follows: A method, system, equipment, and medium for evaluating the multi-scenario regulation capability of an energy storage power station include: acquiring actual operating data of the energy storage power station facing peak shaving, frequency regulation, and emergency power support scenarios; inputting the actual operating data into a pre-constructed comprehensive evaluation model for the multi-scenario regulation capability of energy storage, and transforming the comprehensive evaluation model into a Markov decision process; solving the Markov decision process using a dual deep Q-network algorithm based on an attention mechanism improvement to obtain a comprehensive score for the multi-scenario regulation capability of energy storage; wherein, the comprehensive evaluation model for the multi-scenario regulation capability of energy storage is based on minimizing comprehensive cost and multi-scenario... Maximizing the performance response of the scene is constructed by combining the optimization objective with constraints. This application realizes the collaborative evaluation of multiple scenarios such as peak shaving, frequency regulation and emergency power support by constructing a multi-dimensional evaluation model. By transforming the evaluation model into a Markov decision process, the time-series dependence in energy storage operation can be effectively handled, the evaluation accuracy of multiple scenarios can be improved, dynamic and online evaluation can be realized, and dependence on historical data can be avoided. Through the improved deep reinforcement learning algorithm and multi-objective optimization model, the conflict and uncertainty in complex environment can be effectively handled, the information redundancy problem in high-dimensional state can be effectively alleviated, and the sensitivity to key operating states can be improved. Attached Figure Description

[0033] Figure 1 This is a flowchart of a method for evaluating the multi-scenario regulation capability of an energy storage power station according to this application; Figure 2 This is a schematic diagram of the energy storage peak shaving principle in this application; Figure 3 This is a diagram of the BESS collaborative power output strategy in this application; Figure 4 Here is a flowchart of the Attention-DDQN algorithm in this application; Figure 5 This is a graph showing the average reward outcome for this application. Figure 6 This is a graph showing the average attention weights of this application; Figure 7(a) shows the power deviation in the peak shaving scenario of this application; Figure 7(b) is a power deviation diagram for the frequency modulation scenario of this application; Figure 7(c) is a power deviation diagram for an emergency power support scenario in this application; Figure 8 This is a graph showing the response rate results for this application; Figure 9 This is a graph showing the scoring results for this application; Figure 10 This is a structural diagram of a multi-scenario regulation capability assessment system for an energy storage power station according to this application; Figure 11 This is a diagram of an electronic device according to this application. Detailed Implementation

[0034] This application proposes a method for evaluating the multi-scenario regulation capability of energy storage power stations, aiming to achieve a scientific and accurate assessment of the regulation capability of energy storage systems in multiple scenarios, including peak shaving, frequency regulation, and emergency power support. First, based on the correlation between the characteristics of energy storage power stations and the needs of multiple scenarios, relevant indicators for regulation capability assessment are proposed, and an evaluation indicator system is constructed. Second, based on this system, dynamic weights for scenarios are designed by combining the Analytic Hierarchy Process (AHP) and the entropy weight method, constructing a multi-objective evaluation model with the core objectives of minimizing comprehensive cost and maximizing scenario response. Then, by designing a state space containing multi-scenario demand information and battery state information, and a hierarchical multi-objective reward function constructed based on this evaluation model, a Markov game process for energy storage regulation capability assessment is constructed. Furthermore, addressing the problems of excessively high state space dimensionality and low exploration efficiency in complex energy storage control scenarios, the DQN algorithm is improved through attention mechanism enhancement and a dual-network architecture. To better understand this application, the following description, in conjunction with the accompanying drawings and embodiments, further illustrates the content of this application.

[0035] Example 1: A method for evaluating the multi-scenario regulation capability of an energy storage power station, the specific process of which is as follows: Figure 1 As shown, it includes: Step 1: Obtain actual operating data of the energy storage power station in scenarios of peak shaving, frequency regulation, and emergency power support; Step 2: Input the actual operating data into the pre-constructed comprehensive evaluation model of energy storage multi-scenario regulation capability, and transform the comprehensive evaluation model of energy storage multi-scenario regulation capability into a Markov decision process; Step 3: Solve the Markov decision process using a dual deep Q-network algorithm based on an improved attention mechanism to obtain a comprehensive score of the energy storage multi-scenario regulation capability. The comprehensive evaluation model for energy storage multi-scenario regulation capability is constructed by combining constraints with the optimization objectives of minimizing comprehensive cost and maximizing multi-scenario performance response.

[0036] Before step 1, the process also includes constructing a comprehensive evaluation model for the multi-scenario regulation capability of energy storage. This construction process specifically includes: Based on multi-scenario operation strategies for energy storage, an evaluation index system for the regulation capacity of energy storage power stations is constructed, including: The energy storage multi-scenario operation strategies are as follows: ① Peak Shaving and Frequency Regulation Coordinated Operation: In multi-scenario operation strategies for energy storage, decoupled control based on load curves is a typical multi-timescale coordination method. This strategy divides energy storage into peak shaving and frequency regulation periods by real-time monitoring of grid load. During peak shaving periods, upper and lower limits for peak shaving power are dynamically set: when the load exceeds the upper limit, the energy storage system discharges to reduce peak load; when the load is below the lower limit, it charges to fill valley loads; during non-peak shaving periods, the energy storage switches to participate in frequency regulation to respond to the system's high-frequency power demand, such as... Figure 2 As shown.

[0037]

[0038] In the formula, for Peak-shaving demand power at all times; for Power of load curve at any time; Peak shaving lines are set based on historical load data; Valley filling lines are set based on historical load data.

[0039] ② Emergency Power Support Coordination Operation: In grid operation, in addition to peak shaving and frequency regulation, energy storage systems also need to have the emergency power support capability to cope with sudden power shortages or overloads. Such scenarios typically require energy storage to provide or absorb a large amount of power in a very short time to maintain system stability.

[0040] To achieve coordinated operation of three scenarios—peak shaving, frequency regulation, and emergency power support—this paper proposes a multi-scenario coordination strategy based on priority weight adjustment, such as... Figure 3 As shown.

[0041] During normal operation, energy storage participates in peak shaving and frequency regulation as described above. When the system detects emergency events such as power surges, energy storage immediately switches to emergency power support mode to prioritize responding to system power shortages. After the emergency event ends, energy storage gradually returns to peak shaving or frequency regulation mode based on the current SOC status and system requirements.

[0042] The response priorities for three scenarios are set: emergency power support > frequency regulation > peak regulation; during non-emergency periods, frequency regulation and peak regulation are decoupled on a time scale; when an emergency occurs, the system prioritizes the use of adjustable power resources from energy storage, and may interrupt the current peak regulation or frequency regulation task if necessary.

[0043] Evaluation Indicators: Through the aforementioned multi-scenario coordination strategies, energy storage systems can flexibly respond to grid demands. To further quantify and evaluate its regulation capabilities, a corresponding evaluation indicator system needs to be established. Based on the regulation characteristics of the energy storage system itself, a comprehensive performance evaluation system for energy storage across different application scenarios should be established, considering functionality, safety, and economy.

[0044] The evaluation index system for the regulation capacity of energy storage power stations is as follows: ① Competency assessment indicators: Scene power deviation: Scene power deviation is a core indicator for measuring the regulation capability of an energy storage system. It directly reflects the accuracy of its tracking of power demand in different scenes, i.e., the difference between the demanded power and the actual output power.

[0045]

[0046] In the formula, For power deviation, It supports three different scenarios: peak shaving, frequency regulation, and emergency power. for The charging and discharging power of energy storage required in any given scenario; for The actual charging power released by the stored energy at any given moment.

[0047] Response rate: Defined as the ratio of the number of times a power demand actually responds to the total number of times a demand is requested, reflecting the degree to which an energy storage system responds to various types of demand. A higher response rate indicates that the energy storage system responds to demand more promptly and fully.

[0048]

[0049] In the formula, For response rate, for The number of times a scenario response is effective at any given moment; The number of effective scenario-based response times.

[0050] ② Economic evaluation indicators: Operating costs: Operating costs are mainly used to evaluate the economics of energy storage in actual operation, and consist of the discharge revenue and charging costs of energy storage.

[0051]

[0052] In the formula, Operating costs; for The discharge electricity price at any given moment; for The electricity price for charging at any given moment; for Discharge power at any given moment; for The charging power at any given time.

[0053] Degradation cost: Degradation cost refers to the loss of a battery over long-term operation. This paper uses the rainflow counting method to estimate degradation cost. The core idea is to decompose any charge-discharge process into several complete cycles, and then calculate the degradation cost of each cycle based on the depth of cycle (DOD).

[0054]

[0055] In the formula, The loop depth; It is in a charged state.

[0056] The cycle life of lithium-ion batteries was determined by fitting experimental data. and The relationship is:

[0057] In the formula, This refers to the cycle life of lithium-ion batteries; For cycle life and First coefficient of curve fitting; For cycle life and The second coefficient of the curve fitting; For cycle life and Third coefficient of curve fitting; It is a constant.

[0058] The degradation cost of a complete charge-discharge cycle is defined as:

[0059] In the formula, The degradation cost for a complete charge-discharge cycle; The investment cost per unit capacity of the energy storage system; This refers to the cycle life of a lithium-ion battery.

[0060]

[0061] In the formula, Costs associated with the degradation of energy storage batteries; for The cost of constant degradation; for The cost of constant degradation.

[0062] ③ Safety indicators: State of Charge (SOC): The range of battery state of charge reflects the actual usable energy regulation depth of the system. If the SOC exceeds a reasonable range, it may lead to overcharging or over-discharging of the battery, causing serious safety problems such as internal material aging, accelerated capacity decay, and even thermal runaway. During energy storage discharge, the SOC is:

[0063] During energy storage charging, the SOC is

[0064] In the formula, for The state of charge at any given moment; for The state of charge at any given moment; To be in the time interval The total work performed on the battery at that time; For time intervals; This refers to the battery's rated capacity. For battery discharge efficiency; Improve battery charging efficiency.

[0065] Charging and discharging power: The energy storage charging and discharging power index is a core parameter that measures the electrical energy that an energy storage system can absorb or release per unit time. It directly reflects the system's charging and discharging capabilities and its immediate power support capacity. Strict upper limits need to be set on the charging and discharging power of energy storage systems to ensure that they operate within safe and permissible ranges.

[0066]

[0067]

[0068] In the formula, This refers to the discharge power. This represents the maximum discharge power. for The state of charge at any given moment; It is at its minimum state of charge; This refers to the battery's rated capacity. For battery discharge efficiency; Improve battery charging efficiency; To adjust the duration; This refers to the charging power. This is the maximum charging power; This is the maximum state of charge.

[0069] Based on actual operational data of energy storage power stations under various scenarios over historical periods, and combined with various evaluation indicators, the weights of each evaluation indicator are calculated using the analytic hierarchy process (AHP) and entropy weight method, resulting in the comprehensive weights of the following indicators: (1) Indicator weighting: When evaluating the regulation capacity of energy storage power stations, the weight of the indicators reflects their relative importance, and determining reasonable weights is one of the key issues. This paper adopts an indicator weighting method that combines subjective analytic hierarchy process and objective entropy weighting method. Based on the evaluation indicator system mentioned above, the weights of each regulation capacity evaluation indicator are determined.

[0070] ① Analytic Hierarchy Process (AHP): AHP is a subjective weighting method that uses the 1-9 scale to construct a judgment matrix by comparing pairs of elements and then calculating the eigenvectors of that matrix to determine the weights of each element.

[0071]

[0072] In the formula, For the judgment matrix; This represents the degree of importance of indicator i relative to indicator j at the same level. This represents the number of metrics that need to be compared at this level.

[0073] Calculate the geometric mean of each row of elements and normalize the geometric mean to obtain the final weight.

[0074]

[0075]

[0076] In the formula, The average of all ratios for the j-th indicator; Let be the final weight of the j-th indicator. All weights constitute the weight vector W.

[0077] To ensure consistency in the judgment logic, a consistency check is also required:

[0078]

[0079] In the formula, As a consistency indicator; The average random consistency index is obtained by looking up a table; The consistency ratio; For constructing pairwise comparison judgment matrices; This is the obtained weight vector.

[0080] when If the calculated weights are valid, then the matrix needs to be adjusted until it meets the requirements.

[0081] ② Entropy Weight Method: The entropy weight method is an objective weighting method. During calculation, the original data matrix is ​​first standardized, and then the information entropy and weight of each indicator are calculated.

[0082]

[0083]

[0084] In the formula, This is the standardized value matrix of the i-th indicator under the j-th indicator; Indicators at the same level; The total number of indicators; For indicators at the same level; Let be the value of the i-th indicator under the j-th indicator; Let the information entropy of the j-th indicator be denoted as . Let be the weight of the j-th indicator.

[0085] ③Comprehensive Weighting: Comprehensive weighting combines subjective and objective weightings, and introduces a weighting balance coefficient. Combine weights with complementary properties.

[0086]

[0087] In the formula, For comprehensive weighting; This is the weighting balance coefficient; The objective weights are determined by the entropy weight method. The subjective weights are those used in the analytic hierarchy process.

[0088] Based on the various evaluation metrics and their combined weights, the optimization objectives are to minimize overall cost and maximize performance response across multiple scenarios, including: Based on the established evaluation index system and comprehensive weights, this chapter constructs a multi-objective evaluation model with the core objectives of minimizing comprehensive costs and maximizing performance response across multiple scenarios. This model is used to scientifically evaluate the regulation capabilities of energy storage in various scenarios, such as peak shaving, frequency regulation, and emergency power support. The overall goal of this model is to reasonably balance the system's capability, economy, and safety performance across multiple scenarios. It involves multiple control variables and complex constraints, which are normalized and transformed into the optimization model shown below:

[0089] In the formula, The objective function is... For a specific moment; The total number of time steps represents a 24-hour day and is the total length of a scheduling cycle. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. The overall weighting of safety indicators; For response rate;

[0090] In the formula, For deviation indicators in multiple scenarios; The comprehensive weighting of the deviation in peak shaving scenarios; For power deviation in peak shaving scenarios; The overall weight for frequency modulation scene deviation; For power deviation in frequency modulation scenarios; The overall weighting for deviations in emergency power support scenarios; The overall weight for deviations in emergency power support scenarios.

[0091] The constraints are based on charge / discharge power constraints and state of charge constraints, including: Safety indicators serve as constraints in establishing the evaluation model, ensuring that the indicators remain within the correct range.

[0092] ①Charging and discharging power constraints:

[0093] ② Charge state constraints:

[0094] In the formula, This refers to the discharge power. This represents the minimum discharge power. This represents the maximum discharge power. This refers to the charging power. This is the minimum charging power. This represents the maximum charging power. for The state of charge at any given moment; It is at its minimum state of charge; It is at its maximum state of charge; This represents the initial state of charge at the moment of energy storage. This represents the state of charge at the end of the energy storage process.

[0095] Scoring formula: To scientifically evaluate the comprehensive regulation capabilities of energy storage power stations in various scenarios such as peak shaving, frequency regulation, and emergency power support, this paper proposes a comprehensive scoring function based on the multi-objective evaluation model constructed previously. This function normalizes and sums three categories of indicators: capability, economy, and safety, ultimately obtaining a comprehensive score value between 0 and 100. The higher the score, the stronger the comprehensive regulation capability of the energy storage power station in multiple scenarios.

[0096]

[0097]

[0098] In the formula, The negative impacts of standardizing energy storage power stations; For a specific moment; This represents the total number of time steps. for The number of times a scenario response is effective at any given moment; This is the final positive composite score.

[0099] If the charging / discharging power and SOC exceed the limits during operation in multiple scenarios, the overall score will be directly determined to be 0.

[0100] A comprehensive evaluation model is constructed based on the optimization objective and the constraints.

[0101] Step 1, obtaining actual operational data of the energy storage power station in scenarios of peak shaving, frequency regulation, and emergency power support, specifically includes: Based on the actual operation data obtained by energy storage power stations for peak shaving, frequency regulation and emergency support scenarios.

[0102] Step 2 involves inputting the actual operating data into a pre-constructed comprehensive evaluation model for the multi-scenario regulation capability of energy storage, and transforming the comprehensive evaluation model for the multi-scenario regulation capability of energy storage into a Markov decision process. Specifically, this includes: The evaluation problem of the study can be defined as a Markov chain (MDP), and the MDP model is usually represented by a quintuple:

[0103] In the formula, For state space; For action space; The state transition probability; For the reward function; This is a reward discount factor. Determined by the strategy. Decision in state Select action This forms an MDP chain to describe the interaction between the agent and the environment, and to find a strategy that maximizes the reward.

[0104] ①State Space: Employing a 15-dimensional state vector, it comprehensively characterizes the state of the energy storage power station and its various scenarios. At time... The state is defined as:

[0105] In the formula, for Time-state space; for State of charge at any given moment; For the needs of XX scenario; For the needs of scenario X; For the needs of scenario X; This is historical charge state information; For historical action information; For peak shaving scenario prediction information; For a specific moment; For battery health; This represents the current load level. For energy storage power gradient; The rate of change of state of charge; To meet the comprehensive needs of multiple scenarios; Weights for multi-objective coordination; This refers to the rated capacity. It comprehensively reflects the system operation status and multi-scenario information in battery energy storage power stations.

[0106] ② Action Space: In Markov decision process modeling of reinforcement learning, the action represents the agent at each time step. The actions taken. In the battery energy storage power station described in this article, the actions are the charging and discharging power of the battery:

[0107] In the formula, For action; The charging and discharging power of the battery ③ State transition probability: The probability of transitioning to the next state after taking an action from the current state.

[0108] In the formula, The state transition probability; for Time-state variables; for Time-state variables; for Action variables at time; The probability of a state transition event occurring; for The state value at any given time; for The state value at any given time; for Values ​​are obtained for actions at specific times.

[0109] ④ Reward Function: The objective of evaluating the multi-scenario regulation capability of energy storage power stations is to minimize scenario bias and economic efficiency while maximizing scenario demand response rate. The reward function for a single time step is the reciprocal of the objective function:

[0110] In the formula, For time The reward value; The objective function is... The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; The overall weighting of safety indicators; For response rate.

[0111] The cumulative reward R is the long-term reward that the agent receives based on the impact of each action on the environment. The agent aims to obtain the highest cumulative long-term reward, defined as follows:

[0112] In the formula, For cumulative rewards; This is the discount factor.

[0113] Step 3 involves solving the Markov decision process using a dual deep Q-network algorithm based on an improved attention mechanism to obtain a comprehensive score for the energy storage's multi-scenario regulation capability. Specifically, this includes: An improved DDQN algorithm incorporating an attention mechanism is used to solve the Markov decision process, yielding the optimal charge-discharge power sequence, including: The DQN algorithm uses a deep neural network to estimate the state-action-value function. Because the target network is responsible for both action selection and value evaluation, it can easily lead to overestimation of the Q-value. To address this issue, the DDQN mechanism is introduced, separating action selection and action evaluation. Two independent neural networks perform these tasks respectively: the main network is responsible for action selection, choosing the optimal action from the current state; the target network is responsible for action evaluation, assessing the Q-value of the action selected by the main network. This effectively avoids the estimation bias caused by using the same network for both action selection and target estimation in DQN.

[0114] The update formula for the Q value in DDQN is:

[0115] In the formula, For the goal value; For time steps The reward; Discount factor; Indicates the current Network selection in state The maximum amount that can be received is [number] The value's action is a time step. The state; For time steps The action; for The main network parameters; express The target network parameters; For time steps The state.

[0116] Although the DDQN algorithm effectively alleviates the Q-value overestimation problem by decoupling action selection and value assessment, it still faces challenges in multi-scenario evaluation of energy storage power plants, including redundancy of high-dimensional state information and insufficient processing capabilities for time-series dependencies. To address these issues, this paper improves DDQN by introducing an attention mechanism to dynamically adjust the importance weights of different state features, thereby enhancing the agent's decision-making ability in high-dimensional state spaces.

[0117] An improved DDQN algorithm based on attention mechanism: ①Attention mechanism: During the implementation of the algorithm model, state features are not equally important at each scheduling moment. The features of the agent at a certain scheduling moment may become redundant, thereby interfering with the agent's learning.

[0118] Attention mechanisms are an important technique in deep learning that simulates the human process of allocating attention when processing information. The core idea is to dynamically focus on key information by calculating the importance weight of each input element, thereby improving the ability to extract key features from a high-dimensional state space. In the improved algorithm, the input states are attention-weighted before estimating the Q-value, and the attention-weighted states are then used as inputs to both the main and target networks of DDQN to predict the Q-value. The basic steps are as follows: 1) State space preprocessing Normalize the output of the state space constructed in the Markov game process and map the state values ​​to a specific range.

[0119] 2) Calculation of self-attention weights By calculating the global context vector, a summary of information that can represent the global trend or overall level of the entire state vector is obtained.

[0120]

[0121] In the formula, The global context vector represents the mean characteristic of the state vector, reflecting the overall trend; The first normalized state vector Each element.

[0122] 3) Similarity score calculation Used to measure the correlation between the current specific state and its overall trend.

[0123]

[0124] In the formula, Similarity score; This is the transpose of the normalized state vector.

[0125] 4) Attention weight The similarity scores are converted into a set of weights that are interpretable and usable for weighted convergence, yielding the importance weights of the state vectors in this batch. These weights are then broadcast to the feature dimensions to obtain the attention-weighted states.

[0126]

[0127]

[0128]

[0129] In the formula, The normalized attention weights represent the overall importance percentage of the state. This is the attention weight matrix, used for subsequent weighting. The state after attention weighting highlights important features; Element-wise multiplication of a matrix and a vector; This is the normalized state vector; This is the transpose of the normalized attention weights.

[0130] ② Algorithm Flow: The structure of the improved DDQN algorithm is as follows: Figure 4 As shown, at each time step, after receiving the environmental state, the agent first performs feature weighting through an attention mechanism, and then inputs it into the main network to select an action. The target network is then used to evaluate the value of the action. This improvement effectively alleviates the information redundancy problem in high-dimensional states and enhances the sensitivity to key operational states.

[0131] The values ​​of each evaluation index are obtained by calculating based on the optimal charge and discharge power sequence and various evaluation indicators. The comprehensive score of energy storage's multi-scenario adjustment capability is obtained by substituting the values ​​of each evaluation indicator into the scoring formula.

[0132] The following example uses actual operational data from a month's worth of lithium-ion battery energy storage in a certain region of China for peak shaving, frequency regulation, and emergency power support scenarios to illustrate this application in detail: This section's example demonstrates the use of a lithium-ion battery energy storage power station at the grid end for peak shaving, frequency regulation, and emergency power support. Specific parameters of the lithium-ion energy storage power station are shown in Table 1.

[0133] Data sheets related to lithium-ion battery energy storage power stations The load data used in this example comes from a month's worth of actual operational data of lithium-ion battery energy storage in a certain region of China, covering peak shaving, frequency regulation, and emergency power support scenarios. The sampling period for peak shaving and frequency regulation scenarios is 15 minutes, covering a complete 24-hour operating cycle. Emergency power support sampling data is generated based on historical fault statistics of the regional power grid, divided into five typical emergency event scenarios each day, each lasting one hour. The data is randomly divided into training and test sets; the training set contains 30 days of data, and the test set contains the remaining 1 day's data. In terms of scenarios, peak shaving is relatively simple, frequency regulation tests random response and coordination capabilities, while emergency power support tests the algorithm's priority handling and emergency event perception capabilities.

[0134] The relevant data for peak shaving and frequency modulation are shown in Table 2:

[0135] Peak shaving and frequency modulation related data tables Arbitrage profits from energy storage systems are realized based on the peak-valley electricity prices in the region, as shown in Table 3:

[0136] Algorithm-related data tables (2) Comprehensive Weighting Results: To scientifically evaluate the regulation capability of energy storage power stations under multiple scenarios, the weights of indicators for each scenario were determined based on the constructed evaluation index system. Weighting balance coefficients were set based on expert experience and sample data. .

[0137] In this context, the charging and discharging power and the state of charge (SOC) are used as constraints and are not included in the calculation of the objective function weights.

[0138] The weights of each indicator are shown in Table 4:

[0139] Indicator Weight Table (3) Analysis of evaluation results: ① Analysis of Algorithm Training Results: In order to fully verify the effectiveness and superiority of the proposed improved DDQN algorithm based on attention mechanism, this paper selects DDQN, traditional DQN and genetic algorithm for comparative analysis.

[0140] DDQN algorithm: mainly to verify the effectiveness of the attention mechanism, and to quantitatively evaluate the specific contribution of the attention mechanism to improving the ability to extract state features and improve decision accuracy.

[0141] The DQN algorithm primarily verifies the combined effect of improvements such as the dual-network architecture and priority experience replay, and evaluates the impact of algorithm improvements on convergence, stability, and final performance.

[0142] Genetic Algorithm: This mainly verifies the adaptability of deep reinforcement learning compared to traditional optimization algorithms in complex dynamic environments, and evaluates the technical advantages of deep reinforcement learning in handling temporal dependencies, state awareness, and online learning.

[0143] The average reward curves obtained by training these algorithms in the same energy storage power station environment are as follows: Figure 5 As shown: from Figure 5 As can be seen, at the beginning of training, the average reward curves of the three algorithms, Attention-DDQN, DDQN, and DQN, all exhibited significant oscillations. This is because the agent needs to continuously "trial and error" to learn and update its action policy. As the number of training rounds increases, the agent gradually learns a better policy, and the average reward value of the learned policy continuously increases until it converges to a smaller fluctuation range.

[0144] Attention-DDQN, DDQN, and DQN converged at 1500, 1900, and 2500 rounds, respectively, with Attention-DDQN achieving a higher final convergence value than DDQN and DQN. The Attention-DDQN algorithm exhibited the best performance throughout the training process, boasting the highest average reward value with relatively high fluctuations. This suggests that Attention-DDQN, to some extent, compensates for the potential overestimation of the value function in traditional DDQN, thereby guiding the agent to select better actions.

[0145] The final state-space average attention weight calculation result in the Attention-DDQN algorithm is as follows: Figure 6 As shown: The state space dimensions 1-15 in the diagram correspond to a 15-dimensional state space. The most important factor in calculating the weights is the overall intensity of demand across multiple scenarios, reflecting the overall strength of peak shaving, frequency regulation, and emergency power support needs, directly impacting the agent's scenario switching decisions. Battery health affects the degradation cost and safety of the energy storage infrastructure. The multi-objective coordination weights dynamically adjust the balance between capability and economic indicators. The agent prioritizes this information during decision-making to ensure the energy storage maintains its basic performance and strategy adjustment capabilities across multiple scenarios. Medium-weighted dimensions focus on scenario-related state dimensions, ensuring basic energy storage performance before addressing responsiveness and accuracy tracking across multiple scenarios. Finally, low-weighted dimensions filter redundant information, resulting in faster algorithm convergence. This attention mechanism effectively enhances the agent's ability to extract key features from the high-dimensional state space, giving it a significant advantage in multi-scenario decision-making.

[0146] ② Evaluation Index Results Analysis: Figures 7(a), 7(b), and 7(c) show the deviation reduction effects of four evaluation optimization methods in peak shaving, frequency regulation, and emergency power scenarios of lithium-ion battery energy storage power stations. The horizontal axis represents the operating time, and the vertical axis represents the power demand value. The curve fluctuations reflect the tracking accuracy of different algorithms for scenario power demand. Because a decoupling strategy is used when peak shaving and frequency regulation power demands exist simultaneously, the energy storage power station does not respond to frequency regulation demands when there are peak shaving demands in the frequency regulation scenario. It can be seen from the figures that the agent trained by Attention-DDQN has the best tracking accuracy for scenario power demand in the three scenarios. The percentage reduction in deviation for the three scenarios by different algorithms is shown in Table 5.

[0147] Percentage reduction in deviation table In peak shaving, frequency modulation, and emergency power support scenarios, Attention-DDQN consistently maintains a high percentage reduction in bias with minimal fluctuations across scenarios, demonstrating precise control over power demands across various scenarios and stable tracking of power requirements. Overall, Attention-DDQN exhibits the strongest multi-scenario comprehensive bias control capability and the best stability among all algorithms, with its attention mechanism demonstrating significant efficiency in extracting and adapting features from different scenarios. Without the attention mechanism, the DDQN algorithm achieves a 100% bias reduction in peak shaving scenarios, but only 70.7% in frequency modulation and 95.8% in emergency power support scenarios, showing significant differences in scenario adaptability. In complex scenarios, performance fluctuates noticeably; it only performs well in simple peak shaving scenarios or high-priority scenarios. In the DQN algorithm, single networks are prone to Q-value overestimation and policy instability, making it difficult to consistently output accurate decisions in long-term scenarios, resulting in insufficient bias reduction capabilities in peak shaving and emergency power support scenarios. As for genetic algorithms, because they cannot dynamically capture temporal correlations and sudden changes in the scene, they are difficult to adapt to the dynamic changes in power demand in real time in the complex and high-dimensional evaluation of multiple scenarios of energy storage power stations, resulting in a generally low level of deviation reduction rate.

[0148] For the response rate metric, since only the power demand needs to be met and power accuracy is not required, the value is generally high, primarily reflecting the energy storage's timely response to the scenario. The algorithm training results are as follows... Figure 8 As shown.

[0149] The response rates are shown in Table 6:

[0150] Response rate percentage table Similar to the deviation index, in the peak shaving scenario, the response rates of all four algorithms are at a high level. This is because the time scale of the peak shaving scenario is relatively fixed and the demand pattern is relatively clear, so the requirements for the dynamic response of the algorithm are low, and all types of algorithms can be well adapted. However, in the frequency regulation scenario, the power grid generates high-frequency random power demand, which increases the requirements for the algorithm's real-time status perception and rapid decision-making capabilities, and some algorithms are difficult to adapt to its randomness. In the emergency power scenario, it is necessary to respond to sudden power shortages in a very short time, which places strict requirements on the algorithm's ability to perceive sudden events and prioritize them, and some algorithms are unable to meet the requirements for immediate response.

[0151] Attention-DDQN can respond to scene requirements effectively in all three scenarios, and has significant advantages in state perception and adaptability in multiple scenarios. DDQN and DQN are insufficient in handling complex temporal states and are also insufficient in long-term decision-making, and cannot fully respond to scene requirements. As for traditional optimization algorithms, the response rate is obviously insufficient, and their optimization and evaluation capabilities in high-dimensional, dynamic and multi-scenario scenarios are insufficient, making it difficult to adapt to real-time decision-making requirements.

[0152] The results of the operating cost index and degradation cost index are shown in Table 7. A value less than 0 represents revenue, and a value greater than 0 represents cost.

[0153]

[0154] Economic Indicator Results Data Table Degradation cost is the battery life loss cost estimated based on the rainflow counting method. A higher value indicates more severe capacity decay and lifespan consumption caused by charge-discharge cycles during multi-scenario adjustments. Data shows that the DQN algorithm has the lowest degradation cost because its single-network architecture suffers from overestimation of Q-values, leading to a conservative strategy that reduces charge-discharge frequency. While this reduces battery loss, it sacrifices operational benefits and power deviations in different scenarios. Attention-DDQN and DDQN have similar degradation costs, with Attention-DDQN slightly higher than DQN. This is because both use a dual-network architecture to solve the Q-value overestimation problem, resulting in a more accurate and flexible strategy. They can adjust charge-discharge depth according to dynamic needs such as peak shaving, frequency adjustment, and emergency power support. Although Attention-DDQN incurs a slight increase in cycle loss compared to DQN, it introduces an attention mechanism to focus on key states such as SOC and battery health, avoiding ineffective charge-discharge, thus resulting in a slightly lower degradation cost than DDQN. The genetic algorithm has the highest degradation cost, primarily because it relies on an offline static model and cannot dynamically capture the temporal dependencies of multiple scenarios, accelerating battery capacity decay and leading to the greatest loss. Operating costs are negative due to the existence of emergency power support scenarios, which require energy storage to urgently discharge and support the grid. This often results in negative discharge revenue for these scenarios. The Attention-DDQN algorithm yields the highest operating revenue. Its attention mechanism accurately extracts high-value state information, such as prioritizing discharge during peak electricity price periods and charging during off-peak periods. It also focuses on sudden demands in emergency power support scenarios to obtain subsidies. Furthermore, its dual-network architecture avoids peak-shifting charging and discharging caused by Q-value overestimation, further improving arbitrage profits. The DDQN algorithm yields the second highest revenue. While its dual-network architecture solves the Q-value overestimation problem, it lacks an attention mechanism and cannot accurately capture some short-term high-value scenarios. The DQN algorithm yields the lowest revenue because its conservative strategy misses many peak-valley arbitrage opportunities, such as missing some peak discharge periods or off-peak charging periods, directly impacting the scale of revenue. The genetic algorithm yields the lowest revenue because its static optimization characteristics prevent it from responding to dynamic changes in multiple scenarios in real time, ultimately resulting in a revenue of only 68.7% of that of Attention-DDQN.

[0155] Based on the final total cost, it can be seen that among the four evaluation algorithms for multi-scenario adjustment of lithium-ion battery energy storage power stations, the Attention-DDQN algorithm can better balance battery degradation and operating benefits, and has a greater advantage in the economic evaluation of multi-scenario adjustment of energy storage.

[0156] ③ Evaluation and scoring result analysis: Based on the calculated energy storage index results and scoring formula, the final regulation capability of the lithium-ion battery energy storage power station is evaluated.

[0157] It can be seen that the SOC strictly adhered to the constraints and did not exceed the limits in all algorithms. Furthermore, the power deviation image also shows that the charging and discharging power did not exceed the limits. Therefore, the final evaluation scores of the four algorithms were calculated normally as follows: Figure 9 As shown, the DDQN algorithm based on an improved attention mechanism proposed in this paper achieved the highest score under the designed evaluation model and scoring formula. This indicates that the lithium-ion battery energy storage power station achieved the greatest regulation capability in multiple scenarios under this evaluation method, outperforming other evaluation methods in various indicators. This verifies the effectiveness and superiority of the proposed method in the intelligent evaluation of the regulation capability of lithium-ion battery energy storage power stations in multiple scenarios.

[0158] Example 2: A multi-scenario regulation capability assessment system for energy storage power stations, with the following structure: Figure 10 As shown, it includes: The data acquisition module is used to acquire actual operating data of energy storage power stations in scenarios such as peak shaving, frequency regulation, and emergency power support. The model conversion module is used to input the actual operating data into a pre-built comprehensive evaluation model of the energy storage multi-scenario regulation capability, and to convert the comprehensive evaluation model of the energy storage multi-scenario regulation capability into a Markov decision process. The algorithm solving module is used to solve the Markov decision process using a dual deep Q-network algorithm based on an improved attention mechanism, and obtain a comprehensive score of the energy storage multi-scenario regulation capability. The comprehensive evaluation model for energy storage multi-scenario regulation capability is constructed by combining constraints with the optimization objectives of minimizing comprehensive cost and maximizing multi-scenario performance response.

[0159] It also includes a model building module, which is specifically used for: Based on the actual operation data of energy storage power stations in various scenarios during historical periods, and combined with various evaluation indicators, the weights of each evaluation indicator are calculated using the analytic hierarchy process and the entropy weight method to obtain the comprehensive weight of each evaluation indicator. Based on each evaluation indicator and its comprehensive weight, the optimization objectives are to minimize the overall cost and maximize the performance response in multiple scenarios. The constraints are based on charging and discharging power constraints and state of charge constraints. A comprehensive evaluation model is constructed based on the optimization objective and the constraints. The evaluation metrics include: scenario power deviation, response rate, operating cost, degradation cost, state of charge (SOC), and charge / discharge power.

[0160] The calculation formula for the optimization objective in the model building module is as follows:

[0161] In the formula, The objective function is... For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. The overall weighting of safety indicators; For response rate;

[0162] In the formula, For deviation indicators in multiple scenarios; The comprehensive weighting of the deviation in peak shaving scenarios; For power deviation in peak shaving scenarios; The overall weight for frequency modulation scene deviation; For power deviation in frequency modulation scenarios; The overall weighting for deviations in emergency power support scenarios; The overall weight for deviations in emergency power support scenarios.

[0163] The model transformation module is specifically used for: Based on the actual operating data and the constraints of the comprehensive evaluation model for the multi-scenario regulation capability of energy storage, the state vector is selected to determine the state space; The action space is determined based on the energy storage charging and discharging power. The state transition probability is determined based on the probability of transitioning to the next state after taking an action in the current state; The reward function is determined by the reciprocal of the objective function of the comprehensive evaluation model of energy storage multi-scenario regulation capability. The reward discount factor is determined based on the time scale of peak shaving, frequency regulation, and emergency power support scenarios. The quintuple of a Markov decision process is determined based on the state space, action space, state transition probability, reward function, and reward discount factor.

[0164] The algorithm solution module includes: The Markov decision process solving submodule is used to solve the Markov decision process using a dual deep Q-network algorithm based on an attention mechanism to obtain the optimal charge and discharge power sequence. The index calculation submodule is used to calculate the value of each evaluation index based on the optimal charge and discharge power sequence and each evaluation index. The scoring calculation submodule is used to calculate the comprehensive score of the energy storage's multi-scenario adjustment capability by substituting the values ​​of each evaluation index into the scoring formula.

[0165] The Markov decision process solution submodule includes: The attention weighting unit is used to perform attention mechanism weighting processing on the state space of the Markov decision process to obtain the attention-weighted state. The optimal action selection unit is used to input the attention-weighted state into the main network of the improved dual deep Q network algorithm to select the current optimal action; The optimal charge / discharge power sequence calculation unit is used to evaluate the current optimal action based on the target network of the improved dual-depth Q-network algorithm to obtain the Q value, and update the main network parameters until the algorithm converges, thereby obtaining the optimal strategy and the optimal charge / discharge power sequence.

[0166] The attention weighting unit is specifically used for: The state space of the Markov decision process is normalized to obtain a normalized state vector. The global context vector is calculated based on the normalized feature vector; The similarity score is calculated based on the global context vector. Normalized attention weights are calculated based on similarity scores to obtain normalized attention weights. An attention weight matrix is ​​generated based on the normalized attention weights; The attention-weighted state is obtained by weighting the attention weight matrix and the normalized state vector.

[0167] The calculation formula for the scoring in the scoring calculation submodule is as follows:

[0168]

[0169] In the formula, The negative impacts of standardizing energy storage power stations; For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. for The number of times a scenario response is effective at any given moment; This is the final positive composite score.

[0170] Example 3: like Figure 11 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0171] The processor may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the multi-scenario regulation capability assessment method of an energy storage power station in the above embodiments.

[0172] Example 4 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the multi-scenario regulation capability assessment method for an energy storage power station described in the above embodiments.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0175] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0177] The above are merely embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application shall be included within the scope of the claims of this application pending approval.

Claims

1. A method for evaluating the multi-scenario regulation capability of an energy storage power station, characterized in that, include: Acquire actual operational data of energy storage power stations in scenarios involving peak shaving, frequency regulation, and emergency power support; The actual operating data is input into a pre-constructed comprehensive evaluation model of energy storage multi-scenario regulation capability, and the comprehensive evaluation model of energy storage multi-scenario regulation capability is transformed into a Markov decision process. The Markov decision process is solved using a dual deep Q-network algorithm based on an improved attention mechanism to obtain a comprehensive score of the energy storage multi-scenario regulation capability. The comprehensive evaluation model for energy storage multi-scenario regulation capability is constructed by combining constraints with the optimization objectives of minimizing comprehensive cost and maximizing multi-scenario performance response.

2. The method according to claim 1, characterized in that, The construction process of the comprehensive evaluation model for the multi-scenario regulation capability of energy storage includes: Based on the actual operation data of energy storage power stations in various scenarios during historical periods, and combined with various evaluation indicators, the weights of each evaluation indicator are calculated using the analytic hierarchy process and the entropy weight method to obtain the comprehensive weight of each evaluation indicator. Based on each evaluation indicator and its comprehensive weight, the optimization objectives are to minimize the overall cost and maximize the performance response in multiple scenarios. The constraints are based on charging and discharging power constraints and state of charge constraints. A comprehensive evaluation model is constructed based on the optimization objective and the constraints. The evaluation metrics include: scenario power deviation, response rate, operating cost, degradation cost, state of charge (SOC), and charge / discharge power.

3. The method according to claim 2, characterized in that, The optimization objective is calculated as follows: In the formula, The objective function is... For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. The overall weighting of safety indicators; For response rate; In the formula, For deviation indicators in multiple scenarios; The comprehensive weighting of the deviation in peak shaving scenarios; For power deviation in peak shaving scenarios; The overall weight for frequency modulation scene deviation; For power deviation in frequency modulation scenarios; The overall weighting for deviations in emergency power support scenarios; The overall weight for deviations in emergency power support scenarios.

4. The method according to claim 1, characterized in that, The process of transforming the comprehensive evaluation model of energy storage multi-scenario regulation capability into a Markov decision process includes: Based on the actual operating data and the constraints of the comprehensive evaluation model for the multi-scenario regulation capability of energy storage, the state vector is selected to determine the state space; The action space is determined based on the energy storage charging and discharging power. The state transition probability is determined based on the probability of transitioning to the next state after taking an action in the current state; The reward function is determined by the reciprocal of the objective function of the comprehensive evaluation model of energy storage multi-scenario regulation capability. The reward discount factor is determined based on the time scale of peak shaving, frequency regulation, and emergency power support scenarios. The quintuple of a Markov decision process is determined based on the state space, action space, state transition probability, reward function, and reward discount factor.

5. The method according to claim 1, characterized in that, The Markov decision process is solved using an improved dual deep Q-network algorithm based on an attention mechanism to obtain a comprehensive score of the energy storage's multi-scenario regulation capability, including: The Markov decision process is solved using a dual deep Q-network algorithm based on an improved attention mechanism to obtain the optimal charge-discharge power sequence; The values ​​of each evaluation index are obtained by calculating based on the optimal charge and discharge power sequence and various evaluation indicators. The comprehensive score of energy storage's multi-scenario adjustment capability is obtained by substituting the values ​​of each evaluation indicator into the scoring formula.

6. The method according to claim 5, characterized in that, The method employs an improved dual deep Q-network algorithm based on an attention mechanism to solve the Markov decision process, obtaining the optimal charge-discharge power sequence, including: The state space of the Markov decision process is weighted by an attention mechanism to obtain the attention-weighted state. The attention-weighted state is input into the main network of the improved dual deep Q-network algorithm to select the current optimal action; The target network, based on the improved dual-depth Q-network algorithm, evaluates the current optimal action to obtain the Q value and updates the main network parameters until the algorithm converges, thus obtaining the optimal strategy and the optimal charge / discharge power sequence.

7. The method according to claim 6, characterized in that, The attention-weighted state obtained by performing attention-based weighting on the state space of the Markov decision process includes: The state space of the Markov decision process is normalized to obtain a normalized state vector. The global context vector is calculated based on the normalized feature vector; The similarity score is calculated based on the global context vector. Normalized attention weights are calculated based on similarity scores to obtain normalized attention weights. An attention weight matrix is ​​generated based on the normalized attention weights; The attention-weighted state is obtained by weighting the attention weight matrix and the normalized state vector.

8. The method according to claim 5, characterized in that, The calculation formula for the scoring is as follows: In the formula, The negative impacts of standardizing energy storage power stations; For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. for The number of times a scenario response is effective at any given moment; This is the final positive composite score.

9. A multi-scenario regulation capability assessment system for energy storage power stations, characterized in that, include: The data acquisition module is used to acquire actual operating data of energy storage power stations in scenarios such as peak shaving, frequency regulation, and emergency power support. The model conversion module is used to input the actual operating data into a pre-built comprehensive evaluation model of the energy storage multi-scenario regulation capability, and to convert the comprehensive evaluation model of the energy storage multi-scenario regulation capability into a Markov decision process. The algorithm solving module is used to solve the Markov decision process using a dual deep Q-network algorithm based on an improved attention mechanism, and obtain a comprehensive score of the energy storage multi-scenario regulation capability. The comprehensive evaluation model for energy storage multi-scenario regulation capability is constructed by combining constraints with the optimization objectives of minimizing comprehensive cost and maximizing multi-scenario performance response.

10. The system according to claim 9, characterized in that, It also includes a model building module, which is specifically used for: Based on the actual operation data of energy storage power stations in various scenarios during historical periods, and combined with various evaluation indicators, the weights of each evaluation indicator are calculated using the analytic hierarchy process and the entropy weight method to obtain the comprehensive weight of each evaluation indicator. Based on each evaluation indicator and its comprehensive weight, the optimization objectives are to minimize the overall cost and maximize the performance response in multiple scenarios. The constraints are based on charging and discharging power constraints and state of charge constraints. A comprehensive evaluation model is constructed based on the optimization objective and the constraints. The evaluation metrics include: scenario power deviation, response rate, operating cost, degradation cost, state of charge (SOC), and charge / discharge power.

11. The system according to claim 10, characterized in that, The calculation formula for the optimization objective in the model building module is as follows: In the formula, The objective function is... For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. The overall weighting of safety indicators; For response rate; In the formula, For deviation indicators in multiple scenarios; The comprehensive weighting of the deviation in peak shaving scenarios; For power deviation in peak shaving scenarios; The overall weight for frequency modulation scene deviation; For power deviation in frequency modulation scenarios; The overall weighting for deviations in emergency power support scenarios; The overall weight for deviations in emergency power support scenarios.

12. The system according to claim 9, characterized in that, The model transformation module is specifically used for: Based on the actual operating data and the constraints of the comprehensive evaluation model for the multi-scenario regulation capability of energy storage, the state vector is selected to determine the state space; The action space is determined based on the energy storage charging and discharging power. The state transition probability is determined based on the probability of transitioning to the next state after taking an action in the current state; The reward function is determined by the reciprocal of the objective function of the comprehensive evaluation model of energy storage multi-scenario regulation capability. The reward discount factor is determined based on the time scale of peak shaving, frequency regulation, and emergency power support scenarios. The quintuple of a Markov decision process is determined based on the state space, action space, state transition probability, reward function, and reward discount factor.

13. The system according to claim 9, characterized in that, The algorithm solution module includes: The Markov decision process solving submodule is used to solve the Markov decision process using a dual deep Q-network algorithm based on an attention mechanism to obtain the optimal charge and discharge power sequence. The index calculation submodule is used to calculate the value of each evaluation index based on the optimal charge and discharge power sequence and each evaluation index. The scoring calculation submodule is used to calculate the comprehensive score of the energy storage's multi-scenario adjustment capability by substituting the values ​​of each evaluation index into the scoring formula.

14. The system according to claim 13, characterized in that, The Markov decision process solution submodule includes: The attention weighting unit is used to perform attention mechanism weighting processing on the state space of the Markov decision process to obtain the attention-weighted state. The optimal action selection unit is used to input the attention-weighted state into the main network of the improved dual deep Q network algorithm to select the current optimal action; The optimal charge / discharge power sequence calculation unit is used to evaluate the current optimal action based on the target network of the improved dual-depth Q-network algorithm to obtain the Q value, and update the main network parameters until the algorithm converges, thereby obtaining the optimal strategy and the optimal charge / discharge power sequence.

15. The system according to claim 14, characterized in that, The attention weighting unit is specifically used for: The state space of the Markov decision process is normalized to obtain a normalized state vector. The global context vector is calculated based on the normalized feature vector; The similarity score is calculated based on the global context vector. Normalized attention weights are calculated based on similarity scores to obtain normalized attention weights. An attention weight matrix is ​​generated based on the normalized attention weights; The attention-weighted state is obtained by weighting the attention weight matrix and the normalized state vector.

16. The system according to claim 13, characterized in that, The calculation formula for the scoring in the scoring calculation submodule is as follows: In the formula, The negative impacts of standardizing energy storage power stations; For a specific moment; This represents the total number of time steps. The comprehensive weighting of functional indicators; For deviation indicators in multiple scenarios; This represents the maximum power deviation across multiple scenarios; The comprehensive weighting of economic indicators; Economic indicators for multiple scenarios; This represents the maximum value of the overall cost. for The number of times a scenario response is effective at any given moment; This is the final positive composite score.

17. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a method for evaluating the multi-scenario regulation capability of an energy storage power station as described in any one of claims 1 to 8 is implemented.

18. A readable storage medium, characterized in that, It contains an execution program, which, when executed, implements a method for evaluating the multi-scenario regulation capability of an energy storage power station as described in any one of claims 1 to 8.

Images