An energy storage system power market bidding optimization method and system

Abstract

This invention discloses a method and system for optimizing electricity market bidding for energy storage systems, relating to the field of market bidding optimization technology. The method includes the following steps: S1: Constructing a physical twin layer and a federated sensing layer; mapping the real-time state of energy storage and simulating lifetime decay using digital twin technology; collecting relevant data from the electricity market, energy storage devices, and carbon market; and completing data preprocessing. S2: Based on meta-reinforcement learning and incomplete information game theory, utilizing the preprocessed data from S1, rapidly adjusting the bidding strategy using a small amount of real-time data during the online adaptation phase, and predicting the distribution of opponent strategies. S3: Constructing a three-dimensional optimization objective function of market game theory, lifetime, and carbon efficiency; using the bidding strategy from S2 as the initial solution, solving the multi-objective optimization problem to obtain the optimal bid curve and charging / discharging plan. S4: Executing the optimal bid curve and charging / discharging plan from S3 through edge computing and model predictive control; feeding back to the digital twin model and optimization model to complete dynamic strategy correction and form a closed-loop optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of market bidding optimization, and in particular to an energy storage system power market bidding optimization method and system. Background Art

[0002] At present, the market-oriented reform of the power market has been steadily deepened, the market bidding mechanism has been continuously optimized, the number of entities participating in bidding has been increasing, and the structure has become increasingly diversified. As a key force in the power system undertaking peak and frequency regulation and promoting new energy accommodation, energy storage power stations have an increasingly urgent demand for participating in market bidding to improve revenue. The carbon trading market mechanism is increasingly improved, and the coordinated and linked development of the power market and the carbon market has become the mainstream trend in the industry. At the policy level, it is clearly required that on the basis of exerting their core power regulation function, energy storage power stations should concurrently consider the exploration and realization of low-carbon value, so as to provide support for the clean and low-carbon transformation of the energy structure.

[0003] Traditional bidding methods lack systematic multi-objective consideration and dynamic adaptation mechanism, and often focus on a single objective: they either blindly pursue short-term bidding revenue, have no real-time life data support, and excessive charging and discharging consumes the service life of energy storage equipment, resulting in a sharp increase in equipment operation and maintenance costs, greatly shortened service life, and affecting long-term operating benefits; or they excessively conservatively protect equipment service life and miss market profit opportunities. In addition, most of them do not combine carbon market policies to explore carbon efficiency value, so it is not easy to realize low-carbon revenue monetization, which ultimately leads to low comprehensive benefits of energy storage power stations and makes it difficult to meet the high-quality development requirements of the power market. Summary of the Invention

[0004] The objective of the present invention is to provide an energy storage system power market bidding optimization method and system, which is provided aiming at the disadvantages existing in the prior art that traditional methods often focus on a single objective: they either blindly pursue short-term bidding revenue, and excessive charging and discharging consumes the service life of energy storage equipment, resulting in a sharp increase in equipment operation and maintenance costs, greatly shortened service life, and affecting long-term operating benefits; or they excessively conservatively protect equipment service life and miss market profit opportunities.

[0005] In order to achieve the above objective, the present invention adopts the following technical solution: An energy storage system power market bidding optimization method, comprising the following steps: S1: Construct a physical twin layer and a federal perception layer, map the real-time state of energy storage and simulate life decay through digital twin technology, realize multi-agent data sharing without leaking original data by means of federated learning, collect relevant data of the power market, energy storage equipment and the carbon market, and complete data preprocessing; S2: Based on meta reinforcement learning and incomplete information game theory, use the preprocessed data described in S1 to train a basic dynamic bidding strategy model, quickly adjust the bidding strategy through a small amount of real-time data in the online adaptation stage, and predict the strategy distribution of competitors; S3: Construct a three-dimensional optimization objective function of market game, lifetime and carbon efficiency, take the bidding strategy described in S2 as the initial solution, combine dynamic lifetime cost and carbon benefit, solve the multi-objective optimization problem to obtain the optimal price curve and charge-discharge plan; S4: Through edge computing and model predictive control, execute the optimal price curve and charging / discharging plan described in S3, analyze the market clearing results in real time, and feed them back to the digital twin model and optimization model to complete the dynamic correction of the strategy and form a closed-loop optimization.

[0006] The above technical solution further includes: Specifically, the physical twin layer in S1 includes an energy storage device status acquisition module and a lifetime simulation module. The energy storage device status acquisition module is used to collect the SOC, charge and discharge power, temperature and cycle number of the energy storage battery. The lifetime simulation module simulates the real-time lifetime decay process of the energy storage device based on the collected status data and through a decay model.

[0007] Specifically, the federated perception layer in S1 adopts a horizontal federated architecture in federated learning. The multiple entities include energy storage power stations, power grid companies, electricity sales companies, and carbon trading platforms. Each entity interacts with data through federated nodes, sharing only model parameters and not disclosing its own original data, thus ensuring data privacy and security.

[0008] Specifically, the preprocessed data mentioned in S1 includes data cleaning, data normalization, outlier removal, and data completion. Among them, the electricity market data includes real-time electricity prices, day-ahead electricity price forecasts, market clearing rules, and load forecasts; the energy storage equipment data includes equipment parameters, operating loss data, and lifespan-related data; and the carbon market data includes real-time carbon prices, carbon emission reduction coefficients, and carbon trading rules.

[0009] Specifically, the meta-reinforcement learning described in S2 adopts a model-independent meta-learning framework. The dynamic bidding basic strategy model includes a feature extraction module, a strategy generation module, and a strategy adjustment module. During the online adaptation phase, only a small amount of real-time market data needs to be input to complete the rapid adjustment of the bidding strategy.

[0010] Specifically, the prediction of the opponent's strategy distribution described in S2 adopts the Bayesian estimation method. Based on historical bidding data, past behavior characteristics of the opponent, and real-time market data, it predicts the opponent's bidding range, bidding frequency, and bidding priority, and transforms the prediction results into bidding risk constraints.

[0011] Specifically, the three-dimensional optimization objective function of market game theory, lifespan, and carbon efficiency described in S3 takes maximizing market game benefits, minimizing energy storage lifespan loss, and maximizing carbon benefits as optimization objectives. The priority of the three objectives is adjusted by weighting coefficients. The dynamic lifespan cost is calculated based on the lifespan decay data simulated in S1, and the carbon benefits are calculated based on carbon market data and the carbon emission reduction of energy storage charging and discharging.

[0012] Specifically, the NSGA-Ⅲ algorithm is used to solve the multi-objective optimization problem described in S3. The bidding strategy described in S2 is used as the initial solution. The real-time lifetime decay coefficient and bidding risk constraint are introduced as dynamic boundaries. The Pareto optimal solution set is obtained by iterative solution, and the optimal bid curve and charging and discharging plan are selected from it.

[0013] Specifically, the edge computing described in S4 is deployed at the edge nodes of the energy storage terminal to receive market clearing results and energy storage operation status data in real time, thereby reducing data transmission latency; The model predictive control uses 15 and 30-minute control cycles to track the execution of the charge and discharge plan in real time and correct execution deviations in a timely manner.

[0014] Specifically, the closed-loop optimization process described in S4 is as follows: the market clearing result and the execution deviation data of the charging and discharging plan are synchronously fed back to the digital twin model and the optimization model. The digital twin model updates the simulation results of energy storage lifetime decay, and the optimization model adjusts the parameters of the three-dimensional optimization objective function and the bidding strategy to achieve iterative iteration.

[0015] An energy storage system electricity market bidding optimization system, applied to the method, includes the following steps: S5; Sensing and preprocessing unit, used to collect and process data related to the electricity market, energy storage devices and carbon market; S6; Strategy learning and optimization unit, connected to the perception and preprocessing unit, is used to generate a dynamic bidding strategy based on the data, and obtain the optimal bid curve and charging / discharging plan through multi-objective optimization; S7; Control and execution unit, connected to the strategy learning and optimization unit, is used to execute the optimal price curve and charging and discharging plan, and feed the execution results back to the aforementioned unit to form a closed-loop optimization.

[0016] Specifically, S51; the sensing and preprocessing unit includes a physical twin module for simulating energy storage lifetime and a federated sensing module for secure data collaboration; S61; The strategy learning and optimization unit includes a dynamic bidding strategy model based on meta-reinforcement learning, a Bayesian estimation module for predicting opponents, and a multi-objective optimization solution module for solving three-dimensional optimization objectives; S71; The control and execution unit includes an edge computing module for local fast processing and a model prediction control module for tracking and correcting plans.

[0017] The present invention has the following beneficial effects: This invention achieves synergistic optimization of bidding revenue, energy storage lifespan protection, and carbon efficiency value for energy storage power stations. It effectively addresses the shortcomings of traditional bidding methods in coordinating multiple objectives such as bidding revenue, energy storage lifespan, and carbon efficiency value. By using real-time degradation data obtained through the S1 digital twin lifespan simulation module, dynamic lifespan costs are calculated. Combined with the S3 market game theory and the three-dimensional optimization objective function and weight adjustment mechanism for lifespan and carbon efficiency, the invention avoids excessive consumption of energy storage equipment lifespan in pursuit of short-term bidding revenue. At the same time, it realizes low-carbon value realization through carbon revenue accounting. The S2 opponent strategy prediction reduces bidding risks, and the S4 closed-loop optimization dynamically corrects deviations. This ensures that the bidding scheme is both in line with market realities and can guarantee the long-term stable operation of energy storage equipment, balancing economic benefits, equipment lifespan, and low-carbon benefits. This provides reliable support for energy storage power stations to participate in electricity market bidding in a long-term and efficient manner. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an energy storage system electricity market bidding optimization method and system proposed in this invention; Figure 2 This is a three-dimensional optimization objective diagram of the energy storage system electricity market bidding optimization method proposed in this invention; Figure 3 This is a flowchart of the NSGA-Ⅲ solution proposed in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example

[0021] like Figures 1-3 As shown, the present invention proposes an optimization method for electricity market bidding for energy storage systems, comprising the following steps: S1: Construct a physical twin layer and a federated perception layer. Map the real-time status of energy storage and simulate its lifespan decay through digital twin technology. Achieve multi-entity data sharing without leaking the original data by leveraging federated learning. Collect relevant data from the electricity market, energy storage devices, and carbon market and complete data preprocessing. S2: Based on meta-reinforcement learning and incomplete information game theory, using the preprocessed data from S1, a dynamic bidding basic strategy model is trained. In the online adaptation stage, the bidding strategy is quickly adjusted using a small amount of real-time data to predict the opponent's strategy distribution. S3: Construct a three-dimensional optimization objective function of market game, lifetime and carbon efficiency, take the S2 bidding strategy as the initial solution, combine dynamic lifetime cost and carbon benefit, solve the multi-objective optimization problem to obtain the optimal price curve and charge-discharge plan; S4: Through edge computing and model predictive control, execute the optimal price curve and charging / discharging plan of S3, analyze the market clearing results in real time, and feed them back to the digital twin model and optimization model to complete the dynamic correction of the strategy and form a closed-loop optimization.

[0022] Furthermore, a physical twin layer and a federated sensing layer are constructed to complete multi-source data acquisition and preprocessing. The physical twin layer includes an energy storage device status acquisition module and a life simulation module. The energy storage device status acquisition module collects key operating parameters such as the state of charge (SOC) of the energy storage battery in real time, real-time charge and discharge power, battery body and working environment temperature, and cumulative charge and discharge cycle count. The lifespan simulation module uses the collected real-time status data as input and simulates the real-time lifespan degradation process of energy storage devices through a built-in degradation model, providing data support for subsequent dynamic lifespan cost accounting. The federated perception layer adopts a horizontal federated architecture in federated learning, adapting to multiple entities such as energy storage power stations, power grid companies, electricity sales companies, and carbon trading platforms. Each entity achieves data interaction by deploying federated nodes, sharing only model parameters and not disclosing their own original data, ensuring data privacy and security. On this basis, it collects three types of relevant data: electricity market, energy storage devices, and carbon market. Electricity market data includes real-time electricity prices, day-ahead electricity price forecasts, market clearing rules, and load forecasts. Energy storage device data includes equipment parameters, operating loss data, and lifespan-related data. Carbon market data includes real-time carbon prices, carbon emission reduction coefficients, and carbon trading rules data. Simultaneously, the collected raw data undergoes preprocessing, specifically including data cleaning (removing duplicates and errors), data normalization (unifying data units and magnitudes), outlier removal (screening out unreasonable data caused by collection fluctuations or equipment malfunctions), and data completion (filling in missing key data), forming well-organized and usable preprocessed data to provide a stable data foundation for subsequent steps. Based on meta-reinforcement learning and incomplete information game theory, the incomplete information game in the power market bidding is defined by defining the game strategy space. The participants are the energy storage power station and other bidding entities in the same region. The strategy space is the time-segmented bidding range and feasible set of charging and discharging power allowed by the market rules collected in S1, which corresponds to the bidding and charging / discharging decision variables in S3 and the opponent's bid prediction range estimated by Bayes in S2. The boundary is constrained by the operating parameters of the equipment in S1. An adaptive payoff matrix is ​​constructed. The payoff value is directly calculated using the three-dimensional objective weighted comprehensive payoff formula in S3. The opponent's payoff is fitted based on historical data in S1 and prior information in S2 Bayes. The Bayesian Nash equilibrium solution steps are clarified. A game belief system is constructed based on the S2 Bayes posterior distribution. The payoff matrix and strategy boundary are embedded into the S2 meta-reinforcement learning reward function to iteratively solve for the equilibrium solution set, which is then used as the initial solution boundary of the NSGA-Ⅲ algorithm to select the optimal solution. The bidding strategy and charging / discharging plan are output, which are adapted to the S4 closed-loop optimization process.

[0023] The dynamic bidding basic strategy model is trained and the opponent's strategy distribution is predicted. Meta-reinforcement learning adopts a model-independent meta-learning framework. The corresponding dynamic bidding basic strategy model includes a feature extraction module, a strategy generation module, and a strategy adjustment module. The feature extraction module parses the preprocessed data obtained from S1 and extracts core features affecting bidding, such as electricity price, market rules, energy storage status, and carbon price. The strategy generation module generates a basic bidding strategy adapted to the normal market environment based on these core features. The strategy adjustment module relies on the fast adaptation characteristics of model-independent meta-learning, requiring only a small amount of real-time market data as input during the online adaptation stage. The basic bidding strategy can be quickly adjusted to adapt to real-time market changes without retraining the model. At the same time, a Bayesian estimation method is used to predict the strategy distribution of market competitors. This method takes the preprocessed data of S1 (including historical bidding data), the past actual bidding behavior characteristics of competitors, and real-time market data as inputs. Through Bayesian probabilistic inference, the posterior probability is updated by combining historical prior information and real-time data. The prediction results are then transformed into bidding risk constraints for the subsequent optimization solution in the third step, so that the optimization calculation is consistent with the real market competition environment. A three-dimensional optimization objective function is constructed, and the NSGA-Ⅲ algorithm is used to solve the multi-objective optimization problem to obtain the optimal price curve and charging and discharging plan. First, a three-dimensional optimization objective function of market game, lifetime and carbon efficiency is constructed, with the core optimization objectives of maximizing market game benefits, minimizing energy storage lifetime loss and maximizing carbon benefits. The dynamic lifetime cost is calculated based on the real-time lifetime decay data obtained from the S1 lifetime simulation module, and the carbon benefits are calculated based on the carbon market data of S1 and the carbon emission reduction generated during the energy storage charging and discharging process. Simultaneously through weighting coefficients , , The priority of the three objectives can be adjusted, and the weighting coefficients can be adjusted according to actual operational needs. If priority is given to ensuring market game gains, then the weighting coefficients should be increased. Prioritizing reducing energy storage lifespan loss will increase Prioritizing carbon gains will increase , The multi-objective form of the three-dimensional optimization objective function is as follows:

[0024] The weighted formula is:

[0025] The pricing scheme here refers to the combination of time-of-use pricing and charging / discharging plans formulated by energy storage power stations when participating in market bidding. Specifically, it includes the charging purchase price, the discharging sales price, and the corresponding charging / discharging power arrangement for each time period. When solving multi-objective optimization problems, the NSGA-Ⅲ algorithm is adopted. The dynamic bidding strategy generated by S2 is used as the initial solution for iterative optimization. The real-time lifetime decay coefficient obtained by S1 and the bidding risk constraint transformed by S2 are introduced as dynamic boundary conditions to limit the reasonable solution range. The algorithm carries out multiple rounds of iterative optimization based on the three-dimensional optimization objective function, continuously updates the optimized solution and generates a Pareto optimal solution set that takes into account the balance of the three objectives. Then, the solution with the best comprehensive performance is selected from the solution set, and finally the optimal price curve and charging and discharging plan are determined. The specific process is as follows: the calculated F1, F2, F3 and the set weight coefficients are substituted into the weighting formula to obtain the comprehensive evaluation value of the corresponding price scheme. By continuously adjusting the parameters of the price curve and charging and discharging plan and repeatedly calculating the score, the scheme with the highest score is selected as the optimal result.

[0026] By employing edge computing and model predictive control to execute the optimal solution and complete closed-loop optimization, the following steps are taken. First, the edge computing module is deployed at the edge node of the energy storage terminal. It can receive the clearing results from the electricity market and the operating status data of the energy storage system in real time locally. By processing the data locally, the time consumption caused by remote transmission is avoided, effectively reducing data transmission and processing latency. Simultaneously, the model predictive control module uses fixed control cycles of 15 and 30 minutes. Within each cycle, it tracks and compares the optimal charging and discharging plan determined by S3 with the actual charging and discharging execution of the energy storage system in real time. Once a deviation between actual operation and plan is detected, the operating parameters are immediately adjusted to ensure that the energy storage system accurately follows the optimal charging and discharging plan, guaranteeing the bidding process. Once the solution is successfully implemented, closed-loop optimization is carried out. The specific process is as follows: the market clearing results obtained after the energy storage system actually participates in market bidding, as well as the operational deviation data generated during the execution of the optimal charging and discharging plan, are simultaneously fed back to the digital twin model of S1 and the optimization model of S3. The digital twin model updates the simulation results of energy storage lifespan decay based on the feedback data, while the optimization model adjusts the relevant parameters of the three-dimensional optimization objective function and the bidding strategy generated by S2 based on the feedback market clearing and execution deviation information. The updated simulation results, objective function parameters, and bidding strategy will be used again for the next round of optimization calculation and decision-making, forming a cyclical iteration of data feedback, model update, and strategy optimization to achieve closed-loop optimization throughout the entire process.

[0027] The triggering conditions for each step of this method are as follows: the triggering condition for step S1 is that the energy storage system starts up in real time after power-on. The energy storage device status acquisition module of the physical twin layer collects operating data such as energy storage battery SOC and charging and discharging power at a fixed frequency of 1 minute / time through the industrial standard Modbus-RTU / TCP protocol. The collected raw data and electricity market and carbon market data are cleaned and normalized in a unified JSON format. The horizontal federation architecture of the federation perception layer uses the TLS1.3 encrypted transmission protocol for each node. The aggregation and update of the model parameters for the day are only completed at 24:00 every day. The raw data of each subject is not transmitted throughout the process. Abnormal data is automatically removed by the 3σ criterion and triggers the local interpolation and completion logic. The dynamic bidding basic strategy model in step S2 automatically starts training after the preprocessed dataset is updated. During the online adaptation phase, strategy adjustment is triggered at a fixed cycle of receiving real-time market data every 15 minutes. The feature extraction, strategy generation, and strategy adjustment modules are called sequentially according to serial logic. The Bayesian estimation module runs in parallel with the strategy model. The opponent strategy distribution is updated every 30 minutes. The output bidding risk constraint is passed to the S3 optimization module in a standardized matrix format. When the model inference is abnormal, the fallback mechanism of the stable strategy in the previous cycle is automatically triggered. The multi-objective optimization solution in step S3 is triggered 4 hours before the day-ahead market declaration on each bidding day. It uses the bidding strategy output by S2 as the initial solution and the real-time lifetime decay coefficient and bidding risk constraints passed by S1 as dynamic boundaries. The NSGA-Ⅲ algorithm iteratively calculates according to a fixed process: initial solution loading, constraint boundary verification, objective function calculation, non-dominated sorting, population update, and Pareto optimal solution set generation. The entire process data interaction adopts the standard CSV format. When the solution is abnormal, the population reset and iteration number adaptive adjustment logic are automatically triggered. Finally, the optimal bid curve and charging and discharging plan that conform to the standard format of electricity market declaration are output. The edge computing module in step S4 is deployed on the local edge gateway of the energy storage. It uses the low-latency MQTT protocol to receive market clearing results and energy storage operation data in real time, with data transmission latency controlled within 100ms. The model prediction control module uses a fixed control cycle of 15-30 minutes and executes a serial process of comparing planned and actual values, calculating deviations, correcting operating parameters, and issuing control commands. The closed-loop optimization step is automatically triggered after each control cycle is completed, and the market clearing results and execution deviation data are synchronously fed back to the digital twin model of S1 and the optimization model of S3 in a unified format.

[0028] The physical twin layer in S1 includes an energy storage device status acquisition module and a lifetime simulation module. The energy storage device status acquisition module is used to collect the SOC, charge and discharge power, temperature and cycle number of the energy storage battery. The lifetime simulation module simulates the real-time lifetime decay process of the energy storage device based on the collected status data and through a decay model.

[0029] Furthermore, the energy storage device status acquisition module is specifically used to collect key operating status parameters of the energy storage battery in real time, such as the state of charge (SOC), real-time charging and discharging power, battery body and working environment temperature, and cumulative charge and discharge cycle count, so as to comprehensively obtain real-time operating data of the energy storage device. The life simulation module uses the various real-time status data acquired by the acquisition module as input. Through the built-in energy storage device attenuation model, it simulates and dynamically extrapolates the aging degree and loss rate of the energy storage battery, accurately obtaining the real-time life attenuation results of the energy storage device under the current operating state. This provides a real and reliable real-time data foundation for dynamic life cost accounting in the subsequent optimization steps.

[0030] The federated perception layer in S1 adopts a horizontal federated architecture in federated learning. Multiple entities include energy storage power stations, power grid companies, electricity sales companies, and carbon trading platforms. Each entity interacts with data through federated nodes, sharing only model parameters and not disclosing their own raw data, thus ensuring data privacy and security.

[0031] Furthermore, the federated perception layer adopts a horizontal federated architecture under the federated learning system. This architecture is compatible with various data participants, including energy storage power stations, power grid companies, electricity sales companies, and carbon trading platforms. Each participant deploys corresponding federated nodes to establish a secure data interaction link. During the data collaborative processing and model training process, all participants' original operating data, market data, and transaction data are kept on their local terminals. No original data is transmitted or shared across participants. Only the parameter information generated during model training is transmitted through federated nodes. This not only enables the collaborative utilization of data resources from multiple participants but also avoids the risk of original data leakage from the source, effectively protecting the data privacy and information security of each participant.

[0032] The preprocessed data in S1 includes data cleaning, data normalization, outlier removal, and data completion. Among them, the electricity market data includes real-time electricity prices, day-ahead electricity price forecasts, market clearing rules, and load forecasts; the energy storage equipment data includes equipment parameters, operating loss data, and lifespan-related data; and the carbon market data includes real-time carbon prices, carbon emission reduction coefficients, and carbon trading rules.

[0033] Furthermore, data cleaning removes duplicate and erroneous information from the original data; data normalization unifies data of different units and magnitudes to a standard range, facilitating subsequent calculations; outlier removal filters out unreasonable data caused by collection fluctuations or equipment malfunctions; and data completion replenishes key data missing during the collection process. The processed electricity market data includes real-time electricity prices, day-ahead electricity price forecasts, market clearing rules, and load forecasts; energy storage equipment data includes equipment parameters, operating loss data, and lifespan-related data; and carbon market data includes real-time carbon prices, carbon emission reduction coefficients, and carbon trading rules. Through the above standardization processes, multi-source heterogeneous data are transformed into a regular and usable dataset, providing a stable data foundation for subsequent strategy model training and optimization.

[0034] S2's meta-reinforcement learning adopts a model-independent meta-learning framework. The dynamic bidding basic strategy model includes a feature extraction module, a strategy generation module, and a strategy adjustment module. During the online adaptation phase, only a small amount of real-time market data needs to be input to complete the rapid adjustment of the bidding strategy.

[0035] Furthermore, the feature extraction module analyzes the pre-processed multi-source data to extract core features affecting bidding, such as electricity price, market rules, energy storage status, and carbon price. The strategy generation module generates a basic bidding strategy adapted to the normal market environment based on these core features. The strategy adjustment module relies on the rapid adaptation characteristics of model-independent meta-learning. In the actual online adaptation phase, only a small amount of the latest real-time market data needs to be input to quickly adjust and optimize the basic bidding strategy. The bidding strategy can be made to fit real-time market changes without retraining the model.

[0036] In S2, the Bayesian estimation method is used to predict the opponent's strategy distribution. Based on historical bidding data, past behavior characteristics of the opponent, and real-time market data, the method predicts the opponent's bidding range, bidding frequency, and bidding priority. The prediction results are then transformed into bidding risk constraints for optimization in S3.

[0037] Furthermore, the prediction of the opponent's strategy distribution adopts the Bayesian estimation method. This method uses preprocessed historical bidding data, the actual bidding behavior characteristics of the opponent in the past, and real-time updated market data as input information. Relying on Bayesian probabilistic inference logic, it combines historical prior information and real-time data to update the posterior probability, and makes quantitative probabilistic predictions of the bidding range, bidding frequency, and bidding priority of other market participants. This yields the specific distribution pattern of the opponent's bidding strategy. The prediction result is then transformed into bidding risk constraints that can be used for optimization calculations. These constraints are directly applied to the multi-objective optimization solution process, enabling subsequent optimization calculations to fit the real market competition environment and improve the rationality and security of bidding decisions. The Bayesian estimation method used in this method for predicting the distribution of competitor bidding strategies relies entirely on the established data link and process logic of this method. Its prior distribution is constructed based on the historical bidding data preprocessed in the S1 step and the past bidding behavior characteristics of market participants. For the three types of competitors' bidding range, bidding frequency, and bidding priority that this method clearly predicts, appropriate conjugate prior distributions are adopted respectively. The prior parameters are directly fitted by the statistical characteristics of the corresponding historical data. The likelihood function represents the conditional probability distribution of the competitor's bidding behavior under the current real-time market conditions. It uses real-time market data collected in stage S1, such as electricity prices, load forecasts, and market clearing rules, as conditional variables to characterize the conditional probability of a competitor exhibiting a target bidding behavior under the corresponding market environment. The posterior update strictly follows Bayes' theorem, matching the fixed 15-30 minute cycle of the online strategy adjustment in stage S2 and the closed-loop optimization in stage S4. Each update incorporates the latest market clearing results and actual competitor bidding behavior data, iteratively correcting the prior distribution to obtain the posterior distribution. The output posterior probability result is directly transformed into the bidding risk constraint explicitly defined in this method, used for multi-objective optimization in stage S3, and continuously iteratively optimized with the closed-loop feedback in stage S4. The likelihood function used in the Bayesian estimation in this method employs existing mature and universally applicable techniques in the field of Bayesian statistical inference, including its underlying mathematical form and conditional probability construction logic. It is also a common application in electricity market bidding game analysis for predicting competitor behavior probabilities.

[0038] The three-dimensional optimization objective function of market game theory, lifespan, and carbon efficiency in S3 aims to maximize market game benefits, minimize energy storage lifespan loss, and maximize carbon benefits. The priority of the three objectives is adjusted by weighting coefficients. The dynamic lifespan cost is calculated based on the lifespan decay data from the S1 simulation, and the carbon benefits are calculated based on carbon market data and the carbon emission reduction from energy storage charging and discharging.

[0039] Furthermore, the objective function formula for the three-dimensional optimization of market game theory, lifespan, and carbon efficiency is as follows:

[0040] The weighted formula for adjusting the priority of the three objectives using weighting coefficients is as follows:

[0041] The entire formula is a calculation criterion used to simultaneously optimize the three objectives of market revenue, energy storage lifetime, and carbon efficiency, and is used for subsequent NSGA-Ⅲ algorithm solutions.

[0042] This represents the market game payoff, determined by the bidding strategy, market clearing price, and charging / discharging power. The optimization objective is to maximize this value, hence it's written in the multi-objective formula as follows: The "maximizing profit" is transformed into the "minimizing negative profit" which is easier for the algorithm to solve.

[0043] : Represents the cost of energy storage lifespan loss. Its data comes directly from the real-time lifespan degradation data calculated by the lifespan simulation module in S1. The optimization objective is to minimize the loss, so the formula directly uses . And perform minimization.

[0044] : Represents carbon gain, calculated based on S1's carbon market data (real-time carbon price, carbon emission reduction factor) and the carbon emission reduction corresponding to energy storage charging and discharging. The optimization objective is to maximize it, therefore it is also written as . To adapt to the minimum solution format.

[0045] In the weighted formula , , These are weighting coefficients, all greater than or equal to 0 and summing to 1. They are used to manually adjust the priority of the three objectives in actual optimization: for example, increasing the weighting coefficient if more profit is prioritized. Those who prioritize battery life will increase .

[0046] The carbon revenue of this method refers to the economic benefit realized through carbon market trading of the net carbon emission reduction generated by the energy storage system through charging and discharging activities instead of fossil fuel power generation. The accounting logic is tied to the charging and discharging activities, and all basic data comes from the carbon market data and energy storage charging and discharging operation data collected in the S1 stage of this method. The carbon market data corresponds to the real-time carbon price, carbon emission reduction coefficient, and carbon trading rule data explicitly collected in the S1 stage, and the energy storage charging and discharging operation data corresponds to the real-time operating condition data such as charging and discharging power, operating losses, and charging and discharging capacity collected by the physical twin layer in the S1 stage. No additional data collection items are required, and it is fully compatible with the established data collection process of this method. The net carbon emission reduction calculation corresponds one-to-one with the time period division, charging and discharging power, and effective charging and discharging capacity of the charging and discharging plan output by the S3 link of this method. The calculation scope covers the core operation scenarios of peak shaving, frequency regulation, new energy consumption, and thermal power substitution in the process of energy storage participating in electricity market bidding. The net carbon emission reduction of a single time period is calculated by deducting the carbon emission corresponding to charging loss based on the effective discharge capacity of that time period and the marginal carbon emission factor of the regional power grid. The total carbon emission reduction of the whole cycle is the sum of the calculation values ​​of each time period. The corresponding carbon revenue function F3(x) = total carbon emission reduction × real-time carbon price collected in S1, where x represents the decision variables corresponding to the charging / discharging plan and the price curve, namely core parameters such as charging / discharging period, charging / discharging power, and charging / discharging capacity. This achieves a direct linkage between carbon revenue and charging / discharging behavior, and also forms a linkage and adaptation with the market bidding environment predicted in S2 and the energy storage life loss simulated in S1. This calculation logic is directly applicable to market game theory and the three-dimensional optimization objective function of lifespan and carbon efficiency. To prioritize market game gains, the weighting coefficients corresponding to those gains should be increased. To prioritize reducing battery life loss, increase the weighting coefficient corresponding to life cost. If you want to prioritize increasing carbon gains, increase the weighting coefficient corresponding to carbon gains.

[0047] Score the "Quotation Proposal" using the following formula: Substitute the calculated F1, F2, and F3, along with the predefined weights ω1, ω2, and ω3, into the weighted formula to calculate a comprehensive evaluation value.

[0048] This evaluation score is a comprehensive score for the current set of "price curves + charging and discharging plans".

[0049] Find the solution with the highest score as the optimal result: Within a reasonable price range and charging / discharging range, the parameters of the price curve and charging / discharging plan are continuously adjusted. Each time an adjustment is made, the comprehensive score is recalculated using the formula. Finally, the set of parameters that gives the formula the highest score is selected. This set of parameters is the optimal price curve and optimal charging / discharging plan that S3 ultimately outputs.

[0050] Among them, the bidding scheme refers to the combination of time-of-use bidding and charging and discharging plans formulated by the energy storage power station when participating in the electricity market bidding, based on its own operational needs. Specifically, it includes the charging purchase price, the discharging sales price and the corresponding charging and discharging power arrangement for each time period. It is the core of S3, which is determined by comprehensive scoring and selection through a three-dimensional optimization objective formula. The optimal bidding scheme selected in the end will be directly used to guide the market bidding and actual charging and discharging operation of the energy storage power station.

[0051] The above formula, through the joint calculation of "maximizing revenue, minimizing lifetime loss, and maximizing carbon revenue", outputs a set of optimal pricing curves and charge / discharge plans that can simultaneously balance these three objectives.

[0052] In S3, the NSGA-Ⅲ algorithm is used to solve the multi-objective optimization problem. The S2 bidding strategy is used as the initial solution. The real-time lifetime decay coefficient and bidding risk constraint are introduced as dynamic boundaries. The Pareto optimal solution set is obtained by iterative solution, and the optimal bid curve and charging and discharging plan are selected from it.

[0053] Furthermore, the multi-objective optimization problem refers to the need to simultaneously consider and optimize three mutually restrictive objectives in the context of electricity market bidding: maximizing market game gains, minimizing energy storage battery lifespan degradation, and maximizing carbon gains. It is not possible to optimize any one objective without affecting the other two, and constraints such as real-time lifespan degradation coefficient and bidding risk must also be met. The NSGA-Ⅲ algorithm is used to solve the multi-objective optimization problem. This algorithm directly uses the dynamic bidding strategy generated in S2 as the initial solution for iterative optimization. At the same time, it introduces the real-time lifetime decay coefficient obtained from lifetime simulation in step one and the bidding risk constraint obtained from Bayesian estimation in S2 as dynamic boundary conditions for the optimization iteration process to limit the reasonable solution range. Under this constraint, the algorithm performs multiple rounds of iterative optimization based on the three-dimensional optimization objective function, continuously updates the optimization solution and generates a Pareto optimal solution set that takes into account the balance of multiple objectives. Finally, the solution with the best comprehensive performance is selected from the optimal solution set to determine the optimal bidding curve and charging and discharging plan that meet the requirements of market bidding and energy storage operation.

[0054] In S4, edge computing is deployed at the edge nodes of energy storage terminals to receive market clearing results and energy storage operation status data in real time, thereby reducing data transmission latency. The model predictive control uses 15 and 30-minute control cycles to track the execution of the charge and discharge plan in real time and correct execution deviations in a timely manner.

[0055] Furthermore, the edge computing module is directly deployed on the edge node of the energy storage terminal, which can receive the clearing results issued by the electricity market and the operating status data of the energy storage system itself in real time locally on the energy storage device. By processing the data locally, the time consumption caused by remote data transmission is avoided, and the latency of data transmission and processing is effectively reduced. Meanwhile, the model prediction control module uses a fixed control cycle of 15 to 30 minutes. In each cycle, it tracks and compares the optimal charging and discharging plan determined in step three with the actual charging and discharging execution of the energy storage system in real time. Once a deviation between the actual operation and the plan is detected, the operating parameters will be corrected and adjusted immediately to ensure that the actual operation of the energy storage system can accurately match the optimal charging and discharging plan and ensure the stable implementation of the bidding scheme.

[0056] The specific process of closed-loop optimization in S4 is as follows: the market clearing results and the execution deviation data of the charging and discharging plan are synchronously fed back to the digital twin model and the optimization model. The digital twin model updates the simulation results of energy storage life decay, and the optimization model adjusts the parameters of the three-dimensional optimization objective function and the bidding strategy to achieve iterative iteration.

[0057] Furthermore, the market clearing results obtained after the energy storage system actually participates in market bidding, as well as the operational deviation data generated during the actual execution of the optimal charging and discharging plan, are synchronously fed back to the digital twin model and the optimization model. The digital twin model updates the simulation calculation results of energy storage life decay based on this actual feedback data, while the optimization model adjusts the relevant parameters of the three-dimensional optimization objective function and the bidding strategy based on the market clearing and execution deviation information. The updated simulation results, objective function parameters, and bidding strategies are then used again for the next round of optimization calculation and decision-making, thus forming a cyclical iteration of data feedback, model update, and strategy optimization to achieve complete closed-loop optimization.

[0058] An energy storage system-based electricity market bidding optimization system includes the following steps: S5; Sensing and preprocessing unit, used to collect and process data related to the electricity market, energy storage devices and carbon market; S6; Strategy learning and optimization unit, connected to the perception and preprocessing unit, is used to generate a dynamic bidding strategy based on the data, and obtain the optimal bid curve and charging / discharging plan through multi-objective optimization; S7; Control and execution unit, connected to the strategy learning and optimization unit, is used to execute the optimal price curve and charging and discharging plan, and feed the execution results back to the aforementioned unit to form a closed-loop optimization.

[0059] Furthermore, the sensing and preprocessing unit (S5) first collects multi-source heterogeneous data (such as electricity price, battery state of charge (SOC), and carbon price) from the electricity market, energy storage equipment, and carbon trading platform, and performs preprocessing such as cleaning and normalization to provide high-quality input for decision-making. The strategy learning and optimization unit (S6) receives the processed data, generates a dynamic bidding strategy based on algorithms such as meta-reinforcement learning, and constructs a multi-objective optimization model that takes into account market revenue, equipment lifespan and carbon revenue to solve the problem. Finally, it outputs an optimal day-ahead market price curve and the corresponding charging and discharging plan.

[0060] The control and execution unit (S7) is responsible for implementing the plan, namely submitting bids in the market and controlling the energy storage system to charge and discharge according to the plan. At the same time, it feeds back the market clearing results (whether the bid is successful and the settlement price) and the actual operational deviation data of the energy storage to units S5 and S6 in real time. Unit S5 uses the feedback data to update the equipment status in the digital twin model, while unit S6 uses the feedback results to adjust and optimize the strategy model for the next cycle, thereby realizing the system's self-learning and continuous optimization.

[0061] A specific implementation is as follows: When participating in day-ahead electricity market bidding, the system starts on trading day D-1. Unit S5 collects electricity price forecasts, the current SOC, health status, and carbon price information of the energy storage system on day D. Based on this data, Unit S6 calculates an optimal solution through a model that balances battery degradation and carbon revenue by discharging during peak hours and charging during off-peak hours. Unit S7 submits this bid curve to the trading center before the specified time on day D-1. During actual operation on day D, Unit S7 controls the energy storage system to execute the predetermined charging and discharging plan and feeds back the actual winning bid volume, electricity price, and actual battery operation data to the system. If the actual electricity price deviates from the forecast, Unit S6 will adjust the bidding strategy accordingly; if the actual battery degradation is faster than expected, the life simulation model of Unit S5 will update its parameters, and Unit S6 will allocate a higher weight to the "life degradation" target in subsequent optimizations, thereby adopting a more battery-protective strategy in subsequent bidding.

[0062] S51; The sensing and preprocessing unit includes a physical twin module for simulating energy storage lifetime and a federated sensing module for secure data collaboration. S61; The strategy learning and optimization unit includes a dynamic bidding strategy model based on meta-reinforcement learning, a Bayesian estimation module for predicting opponents, and a multi-objective optimization solution module for solving three-dimensional optimization objectives; S71; The control and execution unit includes an edge computing module for local fast processing and a model prediction control module for tracking and correcting plans.

[0063] Furthermore, the physical twin module in the sensing and preprocessing unit (S51) collects battery SOC, power, and other states in real time and simulates its lifespan degradation, providing dynamic cost data for optimization. Its federated sensing module, without disclosing the original data of all parties, coordinates with multiple data nodes such as energy storage power stations and power grid companies to obtain comprehensive external environmental information. The strategy learning and optimization unit (S61) makes decisions based on this information: its dynamic bidding strategy model generates a basic bidding strategy; the Bayesian estimation module analyzes and predicts competitor behavior, transforming it into risk constraints; the multi-objective optimization solution module solves for the three objectives of market revenue, lifespan loss, and carbon revenue along with the aforementioned risk constraints, outputting the optimal bid and charge / discharge plan. The control and execution unit (S71) is responsible for execution and feedback: its edge computing module is deployed at the terminal to achieve rapid response; the model prediction control module (with a control cycle of 15-30 minutes as described in S4) tracks and corrects plan execution deviations and feeds back market clearing results and operational deviation data to the aforementioned units, driving the digital twin model update and strategy model adjustment.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing electricity market bidding for energy storage systems, characterized in that, Includes the following steps: S1: Construct a physical twin layer and a federated perception layer. Map the real-time status of energy storage and simulate its lifespan decay through digital twin technology. Achieve multi-entity data sharing without leaking the original data by leveraging federated learning. Collect relevant data from the electricity market, energy storage devices, and carbon market and complete data preprocessing. S2: Based on meta-reinforcement learning and incomplete information game theory, using the preprocessed data described in S1, a dynamic bidding basic strategy model is trained. During the online adaptation phase, the bidding strategy is quickly adjusted using a small amount of real-time data to predict the opponent's strategy distribution. S3: Construct a three-dimensional optimization objective function of market game, lifetime and carbon efficiency, take the bidding strategy described in S2 as the initial solution, combine dynamic lifetime cost and carbon benefit, solve the multi-objective optimization problem to obtain the optimal price curve and charge-discharge plan; S4: Through edge computing and model predictive control, execute the optimal price curve and charging / discharging plan described in S3, analyze the market clearing results in real time, and feed them back to the digital twin model and optimization model to complete the dynamic correction of the strategy and form a closed-loop optimization.

2. The method according to claim 1, characterized in that, The physical twin layer in S1 includes an energy storage device status acquisition module and a lifetime simulation module. The energy storage device status acquisition module is used to collect the SOC, charge and discharge power, temperature and cycle number of the energy storage battery. The lifetime simulation module simulates the real-time lifetime decay process of the energy storage device based on the collected status data and through a decay model.

3. The method according to claim 1, characterized in that, The federated perception layer described in S1 adopts a horizontal federated architecture in federated learning. The multiple entities include energy storage power stations, power grid companies, electricity sales companies, and carbon trading platforms. Each entity interacts with data through federated nodes, sharing only model parameters and not disclosing its own original data, thus ensuring data privacy and security.

4. The method according to claim 1, characterized in that, The preprocessed data described in S1 includes data cleaning, data normalization, outlier removal, and data completion. Among them, electricity market data includes real-time electricity prices, day-ahead electricity price forecasts, market clearing rules, and load forecasts; energy storage equipment data includes equipment parameters, operating loss data, and lifespan-related data; and carbon market data includes real-time carbon prices, carbon emission reduction coefficients, and carbon trading rules.

5. The method according to claim 1, characterized in that, The meta-reinforcement learning described in S2 adopts a model-independent meta-learning framework. The dynamic bidding basic strategy model includes a feature extraction module, a strategy generation module, and a strategy adjustment module. During the online adaptation phase, only a small amount of real-time market data needs to be input to complete the rapid adjustment of the bidding strategy.

6. The method according to claim 1, characterized in that, The prediction of the opponent's strategy distribution described in S2 uses the Bayesian estimation method. Based on historical bidding data, past behavior characteristics of the opponent, and real-time market data, it predicts the opponent's bidding range, bidding frequency, and bidding priority, and transforms the prediction results into bidding risk constraints.

7. The method according to claim 1, characterized in that, The objective function for the three-dimensional optimization of market game theory, lifespan, and carbon efficiency described in S3 is as follows: ; With the optimization objectives of maximizing market game gains, minimizing energy storage lifespan losses, and maximizing carbon gains, the priority of these three objectives is adjusted by weighting coefficients. The weighted formula for adjusting the priority of the three objectives by these weighting coefficients is as follows: .

8. The method according to claim 1, characterized in that, The multi-objective optimization problem described in S3 uses the NSGA-Ⅲ algorithm, taking the bidding strategy described in S2 as the initial solution. It introduces the real-time lifetime decay coefficient and bidding risk constraint as dynamic boundaries, and iteratively solves to obtain the Pareto optimal solution set. From this set, the optimal bid curve and charging and discharging plan are selected. The edge computing described in S4 is deployed at the edge nodes of the energy storage terminal to receive market clearing results and energy storage operation status data in real time, thereby reducing data transmission latency. The model predictive control uses 15 and 30-minute control cycles to track the execution of the charge and discharge plan in real time and correct execution deviations in a timely manner.

9. An energy storage system electricity market bidding optimization system, applied to the method of any one of claims 1 to 7, characterized in that, Includes the following steps: S5; Sensing and preprocessing unit, used to collect and process data related to the electricity market, energy storage devices and carbon market; S6; Strategy The learning and optimization unit, connected to the perception and preprocessing unit, is used to generate a dynamic bidding strategy based on the data, and to obtain the optimal bid curve and charging and discharging plan through multi-objective optimization. S7; Control and execution unit, connected to the strategy learning and optimization unit, is used to execute the optimal price curve and charging and discharging plan, and feed the execution results back to the aforementioned unit to form a closed-loop optimization.

10. The system according to claim 9, characterized in that, S51; The sensing and preprocessing unit includes a physical twin module for simulating energy storage lifetime and a federated sensing module for secure data collaboration. S61; The strategy learning and optimization unit includes a dynamic bidding strategy model based on meta-reinforcement learning, a Bayesian estimation module for predicting opponents, and a multi-objective optimization solution module for solving three-dimensional optimization objectives; S71； The control and execution unit includes an edge computing module for local fast processing and a model prediction control module for tracking and correcting plans.

Images