A multi-scenario adaptive light storage cooperative scheduling method and system

Abstract

The present application relates to the technical field of power system, especially to a kind of multi-scenario adaptive light storage collaborative scheduling method and system, method includes through scene identification model classification and grid operation, microgrid isolated network, virtual power plant aggregation scheduling and extreme weather emergency response etc. Scene, based on feature library extraction parameter constructs multi-objective optimization model, uses reinforcement learning engine iteration solution, outputs optimal strategy and sends to energy storage converter and photovoltaic inverter execution, realizes multi-scenario adaptive adaptation, balances scheduling benefit and energy storage life, enhances cross-energy collaborative ability, improves new energy consumption rate.The present application solves the problem that scheduling scene adaptation granularity is insufficient, and energy storage life and benefit balance is poor.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a multi-scenario adaptive photovoltaic-storage collaborative scheduling method and system. Background Technology

[0002] With the continuous increase in the penetration rate of distributed photovoltaic power and the gradual improvement of the power auxiliary service market, photovoltaic power distribution and storage systems face the complex demand for multi-timescale coordinated scheduling. How to maximize the economic benefits of energy storage systems while ensuring the stable operation of the power grid has become the core challenge.

[0003] Common scheduling methods rely on forecast data to make day-ahead or intraday plans. However, the strong randomness of photovoltaic power output and the suddenness of load fluctuations cause significant deviations between the preset scheduling plan and the actual operating conditions. At the same time, existing methods often consider the charging and discharging strategies of energy storage systems separately from the battery health status. In the process of pursuing scheduling benefits, there is a lack of effective constraints on the degradation of energy storage cycle life, which leads to the dual hidden dangers of rapid capacity degradation and insufficient safety margin in long-term operation of energy storage systems. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-scenario adaptive optical-storage collaborative scheduling method and system to effectively solve the technical problems mentioned in the background art.

[0005] One aspect of the present invention provides a multi-scenario adaptive optical-storage collaborative scheduling method, comprising: Acquire photovoltaic power generation prediction data, energy storage system status parameters, and power grid operation status data, and classify and identify the current operation scenario based on a preset scenario recognition model to determine the current scheduling scenario type; Based on the current scheduling scenario type, extract the corresponding scenario scheduling parameter set from the pre-built scenario feature library, and construct a multi-objective optimization scheduling model for the current scenario based on the scenario scheduling parameter set; The photovoltaic power generation prediction data, energy storage system status parameters, power grid operation status data, and the multi-objective optimization scheduling model are input into the reinforcement learning scheduling engine for iterative solution, and the optimal scheduling strategy is output. The optimal scheduling strategy is sent to the energy storage converter and photovoltaic inverter for execution, and the optimal scheduling strategy is modified in real time according to the demand-side response signal to generate the final scheduling instruction.

[0006] In some embodiments, classifying and identifying the current running scenario based on a preset scenario recognition model to determine the current scheduling scenario type includes: The load fluctuation characteristics, photovoltaic output fluctuation characteristics, and grid frequency deviation characteristics are collected from historical operation data, and the feature data are input into a pre-trained scene recognition neural network for pattern matching. Based on the pattern matching results, the matching probability of each scene category is calculated, and the scene type with the highest matching probability is selected as the current scheduling scene type.

[0007] In some embodiments, the method further includes: when the wind speed, temperature and precipitation parameters in the meteorological data are detected to meet the preset extreme weather threshold, automatically triggering an extreme weather emergency response scenario and raising the scheduling priority of the extreme weather emergency response scenario to the highest level.

[0008] In some embodiments, extracting the corresponding scenario scheduling parameter set from a pre-built scenario feature library according to the current scheduling scenario type includes: For the normal grid-connected operation scenario, extract the peak-valley electricity price time period division parameters, demand-side response incentive coefficient, and upper limit value of energy storage charging and discharging power; For microgrid islanded operation scenarios, dynamic model parameters of load resilience coefficient, diesel generator standby power threshold, and islanded frequency stability control bandwidth are extracted; For virtual power plant aggregation and dispatch scenarios, extract the aggregator's dispatch instruction receiving window, the adjustable capacity ratio coefficient of distributed photovoltaic and the ancillary service market clearing price; For extreme weather emergency response scenarios, the emergency power supply guarantee time threshold, the correction coefficient for the upper limit of energy storage discharge depth, and the linkage parameters of disaster early warning level are extracted.

[0009] In some embodiments, constructing a multi-objective optimization scheduling model for the current scenario based on the scenario scheduling parameter set includes: A dual-objective optimization function is constructed, with the net present value of the scheduling revenue of the energy storage system throughout its entire life cycle as the first optimization objective and the equivalent cycle loss of the energy storage battery as the second optimization objective. The constraints of the allowable operating range of energy storage state of charge, the physical limit constraints of energy storage charging and discharging power, the maximum power point tracking constraints of photovoltaics, and the grid frequency regulation reserve capacity constraints are incorporated into the constraint set of the multi-objective optimization scheduling model. The multi-objective optimization scheduling model is solved using a non-dominated sorting genetic algorithm based on Pareto optimal solutions, generating a set of Pareto optimal scheduling solutions, and selecting the optimal scheduling scheme that balances scheduling benefits and energy storage lifetime.

[0010] In some embodiments, the step of inputting the photovoltaic power generation prediction data, energy storage system state parameters, grid operation status data, and the multi-objective optimization scheduling model into a reinforcement learning scheduling engine for iterative solution includes: A reinforcement learning framework is constructed with a scheduling decision-maker as the intelligent agent, and the state space of the intelligent agent is composed of the energy storage charge status, the peak and valley electricity price time period identifier of the grid, the photovoltaic output prediction deviation, and the demand-side response signal status. The action space of the intelligent agent is composed of the energy storage charging and discharging power adjustment action, the photovoltaic inverter power limiting action, and the demand-side response participation action, and the corresponding action is selected according to the current state. Design a composite reward function that integrates scheduling revenue rewards and energy storage lifetime penalties. The scheduling revenue rewards consist of peak-valley arbitrage revenue, ancillary service subsidy revenue, and demand-side response subsidy revenue. The energy storage lifetime penalties consist of the equivalent cycle loss rate and the capacity decay rate. The reinforcement learning scheduling engine is trained offline using a deep deterministic policy gradient algorithm, and an online fine-tuning strategy is used during real-time scheduling to achieve real-time adaptation to battery state of charge, photovoltaic power fluctuations, and grid peak and valley electricity prices.

[0011] In some embodiments, the multi-objective optimization scheduling model further includes a cross-energy collaborative scheduling module, used to access wind power prediction data and gas turbine operating status data to construct a photovoltaic-storage multi-energy collaborative scheduling framework, including: Obtain short-term power forecast data for wind farms, as well as the start-up and shutdown status and upper and lower output limits of gas turbines; To add the sub-objectives of maximizing wind power consumption and minimizing gas turbine operating costs to the objective function of the multi-objective optimization scheduling model; Based on the complementary characteristics of photovoltaic, wind power and gas turbines, the output ratio of each energy source is dynamically allocated. When photovoltaic output is insufficient, gas turbines are used as backup power sources. When photovoltaic output is excessive, energy storage systems are used to absorb the excess power.

[0012] In some embodiments, sending the optimal scheduling strategy to the energy storage converter and the photovoltaic inverter for execution, and modifying the optimal scheduling strategy in real time based on the demand-side response signal, includes: Receive demand-side response call signals issued by the power grid dispatching platform, and parse the regulation capacity demand, response time period and incentive price contained in the signals; Determine whether the current energy storage charge state meets the charge level conditions for demand-side response participation. If it does, generate a scheduling correction instruction to participate in demand-side response and superimpose the scheduling correction instruction onto the optimal scheduling strategy. The actual state of charge of the energy storage system is updated based on the demand-side response execution results, and the updated state of charge is used as the state input for the next scheduling cycle, thereby realizing closed-loop interaction with the power grid scheduling platform.

[0013] In some embodiments, the step of solving the multi-objective optimization scheduling model using a non-dominated sorting genetic algorithm based on Pareto optimal solutions includes: Initialize a population with multiple scheduling schemes; iteratively optimize the population through selection, crossover, and mutation operations, and output a Pareto optimal scheduling solution set after multiple generations of evolution; The scheduling decision system selects the optimal scheduling scheme that balances scheduling benefits and energy storage lifetime from the Pareto optimal scheduling solution set based on actual operational needs.

[0014] Another aspect of the present invention provides a multi-scenario adaptive optical-storage collaborative scheduling system, comprising a data acquisition module, a scene recognition module, a parameter configuration module, a model construction module, a reinforcement learning solution module, and an instruction execution module; wherein: The data acquisition module is used to acquire photovoltaic power generation prediction data, energy storage system status parameters, and grid operation status data; the scene recognition module is used to classify and identify the current operation scene according to a preset scene recognition model to determine the current scheduling scene type; the parameter configuration module is used to extract the corresponding scene scheduling parameter set from a pre-built scene feature library according to the current scheduling scene type; the model building module is used to build a multi-objective optimization scheduling model under the current scene based on the scene scheduling parameter set; the reinforcement learning solution module is used to input the photovoltaic power generation prediction data, energy storage system status parameters, grid operation status data, and the multi-objective optimization scheduling model into the reinforcement learning scheduling engine for iterative solution and output the optimal scheduling strategy; the instruction execution module is used to send the optimal scheduling strategy to the energy storage converter and photovoltaic inverter for execution, and to modify the optimal scheduling strategy in real time according to the demand-side response signal to generate the final scheduling instruction. Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing a scenario recognition mechanism, adaptive adaptation to various scenarios such as normal grid-connected operation, isolated microgrid operation, virtual power plant aggregation dispatch, and emergency response to extreme weather is achieved, solving the problem of insufficient granularity of scenario adaptation in existing dispatching methods; 2. By constructing a dual-objective optimization model that maximizes scheduling revenue and minimizes energy storage lifetime degradation, and using a reinforcement learning scheduling engine for real-time solution, the relationship between scheduling revenue and energy storage asset lifetime is effectively balanced. 3. By accessing complementary energy data such as wind power and gas turbines, a multi-energy collaborative scheduling framework for photovoltaic and energy storage is constructed to enhance cross-energy collaborative scheduling capabilities; 4. By adding a grid interaction response module, demand-side response signals can be accessed, enabling distributed photovoltaic power to participate in grid frequency regulation and peak shaving, thus significantly improving the renewable energy consumption rate. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the multi-scenario adaptive optical-storage collaborative scheduling method of the present invention; Figure 2 This is an architecture diagram of the multi-scenario adaptive optical-storage collaborative scheduling system of the present invention. Detailed Implementation

[0017] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] This invention provides a multi-scenario adaptive photovoltaic-storage collaborative scheduling method and system. First, a data acquisition module acquires photovoltaic power generation prediction data, energy storage system status parameters, and grid operation status data. Then, a scenario recognition module classifies and identifies the current operating scenario based on a preset scenario recognition model to determine the current scheduling scenario type. Next, a parameter configuration module extracts the corresponding scenario scheduling parameter set from a pre-built scenario feature library based on the current scheduling scenario type. A model building module then constructs a multi-objective optimization scheduling model for the current scenario based on the scenario scheduling parameter set. Further, a reinforcement learning solution module inputs the photovoltaic power generation prediction data, energy storage system status parameters, grid operation status data, and the multi-objective optimization scheduling model into a reinforcement learning scheduling engine for iterative solving, outputting the optimal scheduling strategy. Finally, an instruction execution module sends the optimal scheduling strategy to the energy storage converter and photovoltaic inverter for execution, and adjusts the optimal scheduling strategy in real time based on demand-side response signals to generate the final scheduling instruction.

[0019] Example 1 To better understand each module, this embodiment describes each module in detail.

[0020] First, the data acquisition module is responsible for acquiring real-time photovoltaic (PV) power generation forecast data, energy storage system status parameters, and grid operation status data. The PV power generation forecast data includes the predicted active power of the PV power plant for future periods, the prediction confidence interval, and the PV output fluctuation range. The energy storage system status parameters include the current state of charge of the energy storage batteries, current charge / discharge power, historical charge / discharge curves, equivalent cycle count, capacity decay rate, and battery temperature. The grid operation status data includes grid frequency, node voltage, line power flow, peak / valley electricity price period indicators, and grid dispatch instructions.

[0021] The data acquisition module obtains photovoltaic power generation forecast data through the photovoltaic power forecasting system. This system employs an ultra-short-term power forecasting model based on numerical weather prediction, combining historical output data from the photovoltaic power plant with meteorological data to perform rolling forecasts, outputting a photovoltaic active power forecast sequence for the next fifteen minutes to four hours. The temporal resolution of the photovoltaic power forecast data matches the scheduling cycle, ensuring that the forecast data can support the time granularity requirements of scheduling decisions.

[0022] The data acquisition module obtains the energy storage system's status parameters through the energy storage battery management system. The energy storage battery management system collects the battery's terminal voltage, charging / discharging current, battery temperature, and health status parameters in real time, and calculates the current usable capacity and equivalent cycle count of the energy storage battery using an online internal resistance estimation model. It should be noted that the energy storage system's status parameters include the current state of charge, current usable capacity, rated capacity, maximum charging / discharging power limit, battery module balancing status, and the thermal management system's operating status.

[0023] The data acquisition module obtains grid operation status data through the grid energy management system. This data includes grid frequency deviation, tie-line power exchange values, regional control error indicators, peak-valley electricity price time periods, and ancillary service market clearing prices. The data acquisition module establishes a data connection with the grid dispatching platform via a standard communication interface, receiving grid operation status data according to the data format and transmission cycle agreed upon in the grid dispatching protocol.

[0024] Secondly, the scene recognition module classifies and identifies the current operating scene according to a preset scene recognition model. The scene recognition model is built on a deep neural network and adopts a model structure that combines a long short-term memory network with an attention mechanism. The historical training data of the scene recognition model comes from the historical operation records of photovoltaic power plants, energy storage systems, and power grids. The training samples include load fluctuation characteristics, photovoltaic output fluctuation characteristics, and power grid frequency deviation characteristics under different operating scenarios.

[0025] The scene recognition module performs scene classification and recognition processes including data preprocessing, feature extraction, pattern matching, and scene determination.

[0026] During the data preprocessing stage, the scene recognition module standardizes the collected load fluctuation characteristics, photovoltaic output fluctuation characteristics, and grid frequency deviation characteristics to eliminate the influence of different units on the pattern matching results.

[0027] In the feature extraction stage, the scene recognition module uses a convolutional neural network to extract features from the preprocessed feature data, extracting high-dimensional feature representations.

[0028] During the pattern matching stage, the scene recognition module inputs the extracted high-dimensional features into the pre-trained scene recognition neural network for pattern matching and calculates the matching score for each scene category.

[0029] During the scene determination phase, the scene recognition module selects the scene type with the highest matching score as the current scheduling scene type based on the matching score.

[0030] The scene recognition model can identify various scene types, including normal grid-connected operation, microgrid isolated operation, virtual power plant aggregation and dispatch, and extreme weather emergency response. Normal grid-connected operation corresponds to the normal operation state where the photovoltaic power station is connected to the grid and the energy storage system participates in grid dispatch. Microgrid isolated operation corresponds to the special operating state where the microgrid operates independently from the main grid and the energy storage system provides voltage and frequency support as the main or backup power source.

[0031] The virtual power plant aggregation and dispatch scenario corresponds to the collaborative operation state in which distributed photovoltaic resources participate in electricity market transactions and are subject to unified dispatch by the virtual power plant aggregator.

[0032] Extreme weather emergency response scenarios correspond to the emergency operation state in which photovoltaic output drops sharply or grid facilities are damaged under extreme weather conditions such as typhoons, cold waves, and rainstorms, and the energy storage system enters emergency power supply guarantee mode.

[0033] When the scene recognition module detects that wind speed, temperature, and precipitation parameters in the meteorological data meet preset extreme weather thresholds, it automatically triggers an extreme weather emergency response scenario and elevates the scheduling priority of this scenario to the highest level. The meteorological data comes from a meteorological data service interface, and the scene recognition module receives real-time observation data from meteorological stations and numerical weather forecast products. Extreme weather thresholds are set based on the impact of historical extreme weather events on photovoltaic power output and grid operation, including wind speed thresholds, temperature thresholds, and precipitation intensity thresholds. When any meteorological parameter exceeds its corresponding threshold, the scene recognition module immediately switches to the extreme weather emergency response scenario and simultaneously sends a scenario switching notification and priority elevation instruction to the scheduling decision system.

[0034] Next, the parameter configuration module extracts the corresponding scenario scheduling parameter set from the pre-built scenario feature library based on the current scheduling scenario type. The scenario feature library pre-stores scheduling parameter templates for various scenarios, with each scenario type corresponding to a complete set of scheduling parameters. Based on the scenario type identifier output by the scenario recognition module, the parameter configuration module indexes the corresponding scheduling parameter set from the scenario feature library and passes the scheduling parameter set to the model building module.

[0035] For normal grid-connected operation scenarios, the scheduling parameters extracted by the parameter configuration module include peak-valley electricity price time period division parameters, demand-side response incentive coefficients, and upper limits for energy storage charging and discharging power. Peak-valley electricity price time period division parameters define the electricity price level of the grid in each time period, including the start and end times of peak, mid-peak, flat, and low-peak periods. Demand-side response incentive coefficients define the subsidy standards that users can obtain by participating in demand-side response projects, including capacity subsidy unit price and energy subsidy unit price. Upper limits for energy storage charging and discharging power define the maximum allowable charging and discharging power of the energy storage system under grid-connected operation conditions, taking into account grid access capacity limitations and the rated capacity of the energy storage converter.

[0036] For microgrid islanded operation scenarios, the scheduling parameters extracted by the parameter configuration module include dynamic model parameters of load resilience coefficient, diesel generator standby power threshold, and islanded frequency stability control bandwidth.

[0037] The load resilience coefficient dynamic model parameter definition is a matrix of load power elasticity coefficients that vary with voltage and frequency under islanded grid operation conditions, used to evaluate the impact of load fluctuations on system stability.

[0038] The diesel generator standby power threshold defines the proportion of standby capacity that a diesel generator should retain when operating on an isolated grid, ensuring that the power gap can be quickly filled when photovoltaic output drops sharply.

[0039] The isolated grid frequency stability control bandwidth defines the frequency regulation dead zone and regulation gain of the energy storage system when operating in an isolated grid, ensuring that frequency fluctuations are controlled within a safe range.

[0040] For virtual power plant aggregation dispatch scenarios, the parameter configuration module extracts dispatch parameters including the aggregator's dispatch instruction receiving window, the adjustable capacity ratio coefficient of distributed photovoltaic (PV) systems, and the ancillary service market clearing price. The aggregator's dispatch instruction receiving window defines the time period and response latency requirements for the aggregator to issue dispatch instructions to distributed PV resources. The adjustable capacity ratio coefficient of distributed PV systems defines the proportion of upward and downward peak-shaving capacity that distributed PV resources can provide in aggregation dispatch relative to their rated capacity. The ancillary service market clearing price defines the market clearing price for frequency regulation ancillary services and peak-shaving ancillary services, used to calculate dispatch revenue.

[0041] For extreme weather emergency response scenarios, the parameter configuration module extracts scheduling parameters including an emergency power supply guarantee time threshold, an energy storage discharge depth upper limit correction coefficient, and disaster warning level linkage parameters. The emergency power supply guarantee time threshold defines the minimum time that the energy storage system should guarantee continuous power supply to critical loads under extreme weather conditions. The energy storage discharge depth upper limit correction coefficient defines the relaxation ratio for the energy storage discharge depth under extreme weather conditions, allowing the energy storage system to release energy at a greater discharge depth during emergencies. The disaster warning level linkage parameters define the mapping relationship between the disaster warning level and the adjustment range of the scheduling strategy, dynamically adjusting the conservatism of the scheduling strategy according to the warning level.

[0042] Next, the model building module constructs a multi-objective optimization scheduling model for the current scenario based on the scenario scheduling parameter set. The multi-objective optimization scheduling model takes maximizing scheduling revenue and minimizing energy storage cycle lifetime decay as optimization objectives, and energy storage state of charge constraints, power balance constraints, and grid interaction constraints as optimization conditions.

[0043] When building a multi-objective optimization scheduling model, the model building module first configures the weight coefficients of the objective function and the tightness of the constraints according to the current scenario type: For grid-connected and normal operation scenarios, the model building module takes maximizing scheduling benefits as the primary optimization objective and appropriately relaxes the energy storage lifetime constraints. For microgrid islanded operation scenarios, the model building module prioritizes power supply reliability as the primary optimization objective, ensuring power balance and frequency stability. For virtual power plant aggregation and scheduling scenarios, the model building module prioritizes completing the aggregator's scheduling instructions as its primary optimization goal, while also considering scheduling benefits. For extreme weather emergency response scenarios, the model building module prioritizes emergency power supply as its primary optimization objective, aiming to extend the energy storage power supply time to the greatest extent possible.

[0044] The objective function of the multi-objective optimization scheduling model includes a first optimization objective and a second optimization objective. The first optimization objective is the net present value of the scheduling revenue over the entire lifecycle of the energy storage system, comprehensively considering peak-valley arbitrage revenue, ancillary service subsidy revenue, and demand-side response subsidy revenue. Peak-valley arbitrage revenue comes from the price difference between charging the energy storage system during off-peak hours and discharging it during peak hours. Ancillary service subsidy revenue comes from market subsidies obtained by the energy storage system for participating in grid frequency regulation, peak shaving, and other ancillary services. Demand-side response subsidy revenue comes from incentive subsidies obtained by the energy storage system in response to demand-side response calls from the grid dispatch platform. The second optimization objective is the equivalent cycle loss of the energy storage battery, which comprehensively considers the conversion relationship between the actual charge-discharge cycle count and the equivalent full-cycle cycle count of the energy storage battery within the scheduling cycle. The model building module analyzes the charge-discharge curves of the energy storage battery, calculates the loss to battery life for each charge-discharge cycle, and accumulates the total lifespan loss within the scheduling cycle.

[0045] The constraint set of the multi-objective optimization scheduling model includes constraints on the allowable operating range of energy storage state of charge, physical limits on energy storage charging and discharging power, maximum power point tracking constraints of photovoltaics, and grid frequency regulation reserve capacity constraints.

[0046] The energy storage state of charge (SOC) permissible operating range constraint defines the upper and lower limits of the SOC of the energy storage battery, ensuring that the SOC does not exceed the battery's safe operating range. The energy storage charge / discharge power physical limit constraint defines the rated capacity limit of the energy storage converter and the power limit set by the battery management system. The photovoltaic maximum power point tracking (MPPT) constraint defines that the photovoltaic inverter should operate in MPT mode as much as possible, unless the dispatch strategy requires power limiting. The grid frequency regulation reserve capacity constraint defines the proportion of reserve capacity that the energy storage system should maintain for primary frequency regulation.

[0047] The model building module employs a non-dominated sorting genetic algorithm based on Pareto optimal solutions to solve the multi-objective optimization scheduling model. As can be understood, the non-dominated sorting genetic algorithm is a multi-objective optimization algorithm that searches for a uniformly distributed optimal solution set on the Pareto optimal front through non-dominated sorting and crowding distance calculation. The model building module first initializes a population containing multiple scheduling schemes, then iteratively optimizes the population through selection, crossover, and mutation operations, outputting a Pareto optimal scheduling solution set after multiple generations of evolution. Each solution in the Pareto optimal scheduling solution set satisfies the constraints, and no other solution is simultaneously superior to this solution in all optimization objectives. The scheduling decision system selects the optimal scheduling scheme from the Pareto optimal scheduling solution set, balancing scheduling benefits and energy storage lifetime, based on actual operational requirements.

[0048] Then, the reinforcement learning solution module inputs photovoltaic power generation prediction data, energy storage system state parameters, grid operation status data, and a multi-objective optimization scheduling model into the reinforcement learning scheduling engine for iterative solution, outputting the optimal scheduling strategy. The reinforcement learning scheduling engine dynamically adjusts its action strategy based on real-time perceived changes in energy storage state of charge, photovoltaic power fluctuations, and grid peak-valley electricity prices, reducing energy storage cyclic degradation while ensuring power supply stability.

[0049] The reinforcement learning scheduling engine employs a reinforcement learning framework with a scheduling decision-maker as the agent. Through continuous interaction with the scheduling environment, the agent learns the optimal action to take in different states to maximize long-term cumulative rewards. The training process of the reinforcement learning scheduling engine includes two phases: offline training and online fine-tuning.

[0050] In terms of state space construction, the reinforcement learning scheduling engine uses the energy storage state of charge, grid peak-valley price period identifiers, photovoltaic output prediction deviation, and demand-side response signal status to construct the agent's state space. Specifically, the energy storage state of charge represents the ratio of the current remaining charge of the energy storage battery to its rated capacity, reflecting the energy reserve level of the energy storage system. The grid peak-valley price period identifier indicates the current price period category, including peak, mid-peak, flat, and low-peak periods. The photovoltaic output prediction deviation indicates the degree of deviation between the actual and predicted photovoltaic output, reflecting the severity of photovoltaic power fluctuations. The demand-side response signal status indicates whether the grid dispatch platform has issued a demand-side response call and the specific parameters of the call.

[0051] In terms of action space construction, the reinforcement learning scheduling engine constructs the agent's action space by combining energy storage charging and discharging power adjustment actions, photovoltaic inverter power limiting actions, and demand-side response participation actions. Energy storage charging and discharging power adjustment actions define the power commands of the energy storage system, including charging power commands, discharging power commands, and standby commands. Photovoltaic inverter power limiting actions define the upper limit of the active power of the photovoltaic inverter, used to limit the grid-connected power of photovoltaic power generation when photovoltaic output is excessive. Demand-side response participation actions define whether the energy storage system responds to demand-side response calls and the size of the response capacity.

[0052] In terms of reward function design, the reinforcement learning scheduling engine employs a composite reward function that integrates scheduling revenue rewards and energy storage lifetime penalties. The scheduling revenue rewards consist of peak-valley arbitrage revenue, ancillary service subsidy revenue, and demand-side response subsidy revenue. Peak-valley arbitrage revenue is calculated based on the energy storage system's charging and discharging behavior during the current period and the grid electricity price level. Ancillary service subsidy revenue is calculated based on the frequency regulation and peak shaving ancillary service capacity and duration provided by the energy storage system. Demand-side response subsidy revenue is calculated based on the actual regulation capacity completed by the energy storage system in response to demand-side response calls and the incentive price.

[0053] The energy storage lifetime penalty is composed of the equivalent cycle count loss rate and the capacity decay rate. The equivalent cycle count loss rate reflects the degree to which the current scheduling action consumes the cycle life of the energy storage battery; actions with greater depth of discharge and higher charge / discharge rates correspond to a higher loss rate. The capacity decay rate reflects the impact of the current scheduling action on the degradation of the usable capacity of the energy storage battery. The reinforcement learning scheduling engine incorporates the energy storage lifetime penalty into the total reward function as a negative reward, guiding the agent to consider energy storage lifetime while pursuing scheduling gains.

[0054] The reinforcement learning scheduling engine employs a deep deterministic policy gradient algorithm (DPR) for policy learning. DPR is a deep reinforcement learning algorithm suitable for continuous action spaces, consisting of an actor network and a critic network. The actor network is responsible for outputting deterministic action decisions based on the current state, while the critic network is responsible for evaluating the value of the current state-action pair. The reinforcement learning scheduling engine first trains the DPR using historical running data offline, enabling the agent to learn a basic scheduling policy. After offline training, the engine employs an online fine-tuning strategy during real-time scheduling, making small adjustments to network parameters based on real-time perceived environmental changes to achieve real-time adaptation to battery state of charge, photovoltaic power fluctuations, and peak-valley electricity prices.

[0055] The reinforcement learning solution module outputs the optimal scheduling strategy, which is then passed to the instruction execution module. The optimal scheduling strategy includes the target charging and discharging power command for the energy storage system, the target grid-connected power limit for the photovoltaic inverter, and the participation command for demand-side response. The optimal scheduling strategy is generated based on the current state and the solution results of the multi-objective optimization scheduling model, while also incorporating scheduling experience accumulated by the reinforcement learning scheduling engine through long-term learning.

[0056] Next, the instruction execution module sends the optimal scheduling strategy to the energy storage converter and the photovoltaic inverter for execution. The energy storage converter receives the energy storage charging and discharging power instructions from the optimal scheduling strategy, converts DC power to AC power or reverses it, and drives the energy storage battery to perform charging and discharging operations according to the instructions. The photovoltaic inverter receives the photovoltaic grid-connected power limit instructions from the optimal scheduling strategy, and activates the power limiting function when the photovoltaic output exceeds the limit to ensure that the grid-connected photovoltaic power generation does not exceed the instruction limit.

[0057] During the execution of the optimal scheduling strategy, the instruction execution module monitors the operating status of the energy storage system and the photovoltaic system in real time. When abnormal situations such as the energy storage state of charge approaching the constraint boundary, the power of the energy storage converter exceeding the limit, or the photovoltaic inverter malfunction are detected, the instruction execution module immediately triggers the abnormal handling process, reports the abnormal information to the scheduling decision system, and automatically adjusts the scheduling instructions according to the preset abnormal handling rules.

[0058] The instruction execution module adjusts the optimal scheduling strategy in real time based on demand-side response signals. It receives demand-side response call signals from the power grid dispatching platform via a communication interface. These signals include parameters such as regulating capacity demand, response time period, and incentive price. The instruction execution module parses the content of the demand-side response call signals to extract the regulating capacity demand and response time period information.

[0059] The instruction execution module determines whether the current state of charge (SBC) of the energy storage meets the charge level conditions for demand-side response participation. Demand-side response participation requires the energy storage system to have sufficient available capacity and discharge capability. Based on the current SBC and available charge / discharge power, the instruction execution module calculates the maximum regulation capacity that the energy storage system can respond to. If the maximum adjustment capacity is greater than or equal to the adjustment capacity demand in the demand-side response call signal, then the participation condition is determined to be met. If the maximum adjustment capacity is less than the adjustment capacity requirement, then the participation condition is not met.

[0060] When the participation conditions are met, the instruction execution module generates a scheduling correction instruction for participating in the demand-side response. This instruction is superimposed on the optimal scheduling strategy, increasing or decreasing the energy storage discharge power based on the original scheduling instruction to respond to the demand-side response call from the grid dispatching platform. The superposition of the scheduling correction instruction follows a constraint verification process to ensure that the superimposed scheduling instruction does not lead to violations of energy storage state of charge limits, power limits, or other safety constraints.

[0061] The instruction execution module updates the actual state of charge (SOC) of the energy storage system based on the demand-side response execution results. After the demand-side response is executed, the SOC of the energy storage system changes. The instruction execution module uses this changed SOC as the state input for the next scheduling cycle and passes it to the scenario recognition module and the reinforcement learning solution module. The updated state information enables the scheduling decision system to perform the next round of scheduling optimization based on the latest energy storage system state, achieving closed-loop interaction with the power grid scheduling platform.

[0062] In some possible implementations, the multi-objective optimization scheduling model also includes a cross-energy collaborative scheduling module, which is used to access wind power prediction data and gas turbine operating status data to build a photovoltaic-storage multi-energy collaborative scheduling framework and improve cross-energy collaborative scheduling capabilities.

[0063] The cross-energy collaborative scheduling module acquires short-term power forecast data from wind farms. The wind power forecast data comes from the wind farm power forecast system and includes the predicted active power of wind power for future periods and the prediction error range. The time resolution of the wind power forecast data is consistent with that of the photovoltaic power forecast data, which facilitates multi-energy complementary coordinated scheduling.

[0064] The cross-energy collaborative scheduling module acquires the operating status data of the gas turbine, including its current start / stop status, current output level, rated output upper and lower limits, minimum start / stop time limits, ramp rate limits, and operating cost parameters. The module establishes a data connection with the gas turbine control system to obtain the gas turbine's operating status and scheduling instructions in real time.

[0065] In the objective function of the multi-objective optimization scheduling model, the cross-energy collaborative scheduling module adds two sub-objectives: maximizing wind power consumption and minimizing gas turbine operating costs. The wind power consumption maximization sub-objective encourages scheduling strategies to prioritize wind power consumption and reduce wind curtailment. The gas turbine operating cost minimization sub-objective encourages scheduling strategies to rationally arrange gas turbine start-up, shutdown, and output, reducing fuel consumption and start-up costs.

[0066] The cross-energy coordinated dispatch module dynamically allocates the output ratio of each energy source based on the complementary characteristics of photovoltaic (PV), wind power, and gas turbines. When PV output is insufficient, the module prioritizes the use of gas turbines as backup power, determining the target output of the gas turbines based on load demand and their ramp-up capabilities. When PV output is excessive, the module prioritizes utilizing energy storage systems to absorb the excess PV power. Fluctuations in wind power output are mitigated through the charging and discharging regulation of the energy storage system and adjustments to the output of the gas turbines.

[0067] The scheduling strategy of the photovoltaic-storage multi-energy coordinated scheduling framework comprehensively considers the output characteristics, operating costs and environmental benefits of each energy source. The scheduling decision system optimizes the output arrangement of each energy source and the charging and discharging plan of the energy storage system based on load forecasting and renewable energy output forecasting, with the goal of minimizing system operating costs and maximizing the renewable energy absorption rate.

[0068] Example 2 This embodiment takes the normal grid-connected operation scenario as an example to describe in detail the application process in daily scheduling and operation.

[0069] Under normal grid-connected operation, the data acquisition module continuously collects photovoltaic power generation forecast data, energy storage system status parameters, and grid operation status data. The photovoltaic power forecasting system provides rolling forecasts of the photovoltaic output sequence for the next four hours, outputting forecast values ​​every fifteen minutes. The energy storage battery management system monitors the state of charge, charging and discharging power, battery temperature, and other status parameters of the energy storage batteries in real time and reports these parameters to the dispatch decision system. The grid energy management system publishes real-time operational information such as grid frequency, node voltage, and peak-valley electricity price periods.

[0070] The scene recognition module receives standard data output from the data acquisition module and extracts load fluctuation characteristics, photovoltaic output fluctuation characteristics, and grid frequency deviation characteristics. Load fluctuation characteristics include the load change rate, peak load time, and valley load occurrence times over the past hour. Photovoltaic output fluctuation characteristics include the photovoltaic output change rate and photovoltaic output prediction deviation distribution over the past hour. Grid frequency deviation characteristics include the average, maximum, and minimum frequency deviation values ​​over the past hour. The scene recognition module inputs the extracted feature data into the scene recognition model. The model outputs that the current scene type is a grid-connected normal operation scene with a matching probability of 95%.

[0071] The parameter configuration module receives the identifier of a grid-connected normal operation scenario and extracts the corresponding scheduling parameter set from the scenario feature library. The extracted scheduling parameters include peak-valley electricity price time period division parameters: peak hours are 10:00 AM to 12:00 PM and 6:00 PM to 8:00 PM; mid-peak hours are 12:00 PM to 6:00 PM; flat hours are 7:00 AM to 10:00 AM and 8:00 PM to 10:00 PM; and off-peak hours are 10:00 PM to 7:00 AM the following morning. The extracted scheduling parameters also include demand-side response incentive coefficients: capacity subsidy unit price is 100 yuan per kilowatt, and energy subsidy unit price is 2 yuan per kilowatt-hour. The extracted scheduling parameters also include the upper limit of energy storage charging and discharging power: the upper limit of charging power is the rated power of the energy storage converter, and the upper limit of discharging power is the rated power of the energy storage converter.

[0072] The model building module receives the scheduling parameter set and constructs a multi-objective optimization scheduling model under normal grid-connected operation scenarios. The module configures the objective function weights: the weight for maximizing scheduling revenue (first optimization objective) is set to 0.8, and the weight for minimizing energy storage lifetime degradation (second optimization objective) is set to 0.2. The module also configures constraints: the allowable operating range of energy storage state of charge is set to 20% to 95%; the energy storage charging and discharging power limit is set to the rated power of the energy storage converter; the photovoltaic maximum power point tracking constraint is set to track the maximum power point as much as possible; and the grid frequency regulation reserve capacity is set to 5% of the rated capacity of the energy storage. The module uses a non-dominated sorting genetic algorithm to solve the multi-objective optimization scheduling model, generating a Pareto optimal scheduling solution set.

[0073] The reinforcement learning solution module inputs real-time data and a multi-objective optimization scheduling model into the reinforcement learning scheduling engine. The engine selects the optimal action based on the current energy storage state of charge, grid peak-valley pricing periods, photovoltaic output prediction deviation, and demand-side response signal status. Assuming the current time is 9:30 AM, the energy storage state of charge is 70%, the grid peak-valley pricing period is a flat period, the photovoltaic output prediction deviation is zero, and the demand-side response signal status is "no call." The reinforcement learning scheduling engine selects a standby action for the energy storage based on this status information, maintaining the current energy storage charging and discharging state.

[0074] The reinforcement learning solution module outputs the optimal scheduling strategy, including setting the energy storage charging / discharging power command to 0 and the photovoltaic grid-connected power limit command to unlimited. The instruction execution module sends the optimal scheduling strategy to the energy storage converter and the photovoltaic inverter for execution. The energy storage converter remains in standby mode, and the energy storage battery does not charge or discharge. The photovoltaic inverter operates in maximum power point tracking mode, and all photovoltaic power generation is connected to the grid.

[0075] At 10:00 AM, the peak-valley electricity price period switched to a peak period. The data acquisition module detected the price change and sent the new peak-valley price period identifier to the reinforcement learning solution module. The reinforcement learning solution module updated its state information and reselected actions based on the new information. The reinforcement learning scheduling engine, based on the conditions that the energy storage's state of charge is 70% and the peak-valley price period is a peak period, selected an energy storage discharge action, setting the discharge power to 80% of the energy storage converter's rated power. The reinforcement learning solution module output a new optimal scheduling strategy, and the instruction execution module issued the scheduling instructions to the energy storage converter for execution. The energy storage converter discharged at 80% of its rated power according to the instructions, releasing energy back to the grid and obtaining the discharge benefits during the peak period.

[0076] At 1 PM, the power grid dispatch platform issued a demand-side response call signal, with a call capacity of 20% of the energy storage's rated capacity, a response period from 2 PM to 4 PM, and an incentive price of 3 yuan per kilowatt-hour. The instruction execution module received the demand-side response call signal, determined that the current energy storage's state of charge (SOC) was 45%, meeting the participation conditions. The instruction execution module generated a dispatch correction instruction, increasing the energy storage discharge power by 20% of the rated capacity based on the original optimal dispatch strategy. The instruction execution module then superimposed the dispatch correction instruction onto the optimal dispatch strategy to generate the final dispatch instruction. The energy storage converter executed the final dispatch instruction, increasing its discharge power between 2 PM and 4 PM to respond to the power grid dispatch platform's demand-side response call. After the demand-side response was completed, the instruction execution module updated the actual SOC of the energy storage system and used the updated SOC as the state input for the next dispatch cycle.

[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0078] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-scenario adaptive optical-storage collaborative scheduling method, characterized in that, include: Acquire photovoltaic power generation prediction data, energy storage system status parameters, and power grid operation status data, and classify and identify the current operation scenario based on a preset scenario recognition model to determine the current scheduling scenario type; Based on the current scheduling scenario type, extract the corresponding scenario scheduling parameter set from the pre-built scenario feature library, and construct a multi-objective optimization scheduling model for the current scenario based on the scenario scheduling parameter set; The photovoltaic power generation prediction data, energy storage system status parameters, power grid operation status data, and the multi-objective optimization scheduling model are input into the reinforcement learning scheduling engine for iterative solution, and the optimal scheduling strategy is output. The optimal scheduling strategy is sent to the energy storage converter and photovoltaic inverter for execution, and the optimal scheduling strategy is modified in real time according to the demand-side response signal to generate the final scheduling instruction.

2. The method according to claim 1, characterized in that, The step of classifying and identifying the current operating scenario based on a preset scenario recognition model to determine the current scheduling scenario type includes: The load fluctuation characteristics, photovoltaic output fluctuation characteristics, and grid frequency deviation characteristics are collected from historical operation data, and the feature data are input into a pre-trained scene recognition neural network for pattern matching. Based on the pattern matching results, the matching probability of each scene category is calculated, and the scene type with the highest matching probability is selected as the current scheduling scene type.

3. The method according to claim 2, characterized in that, Also includes: When the wind speed, temperature, and precipitation parameters in the meteorological data are detected to meet the preset extreme weather thresholds, the extreme weather emergency response scenario is automatically triggered, and the scheduling priority of the extreme weather emergency response scenario is raised to the highest level.

4. The method according to claim 1, characterized in that, The step of extracting the corresponding scenario scheduling parameter set from a pre-built scenario feature library based on the current scheduling scenario type includes: For the normal grid-connected operation scenario, extract the peak-valley electricity price time period division parameters, demand-side response incentive coefficient, and upper limit value of energy storage charging and discharging power; For microgrid islanded operation scenarios, dynamic model parameters of load resilience coefficient, diesel generator standby power threshold, and islanded frequency stability control bandwidth are extracted; For virtual power plant aggregation and dispatch scenarios, extract the aggregator's dispatch instruction receiving window, the adjustable capacity ratio coefficient of distributed photovoltaic and the ancillary service market clearing price; For extreme weather emergency response scenarios, the emergency power supply guarantee time threshold, the correction coefficient for the upper limit of energy storage discharge depth, and the linkage parameters of disaster early warning level are extracted.

5. The method according to claim 1, characterized in that, The construction of a multi-objective optimization scheduling model for the current scenario based on the scenario scheduling parameter set includes: A dual-objective optimization function is constructed, with the net present value of the scheduling revenue of the energy storage system throughout its entire life cycle as the first optimization objective and the equivalent cycle loss of the energy storage battery as the second optimization objective. The constraints of the allowable operating range of energy storage state of charge, the physical limit constraints of energy storage charging and discharging power, the maximum power point tracking constraints of photovoltaics, and the grid frequency regulation reserve capacity constraints are incorporated into the constraint set of the multi-objective optimization scheduling model. The multi-objective optimization scheduling model is solved using a non-dominated sorting genetic algorithm based on Pareto optimal solutions, generating a set of Pareto optimal scheduling solutions, and selecting the optimal scheduling scheme that balances scheduling benefits and energy storage lifetime.

6. The method according to claim 1, characterized in that, The step of inputting the photovoltaic power generation prediction data, energy storage system state parameters, power grid operation status data, and the multi-objective optimization scheduling model into the reinforcement learning scheduling engine for iterative solution includes: A reinforcement learning framework is constructed with a scheduling decision-maker as the intelligent agent, and the state space of the intelligent agent is composed of the energy storage charge status, the peak and valley electricity price time period identifier of the grid, the photovoltaic output prediction deviation, and the demand-side response signal status. The action space of the intelligent agent is composed of the energy storage charging and discharging power adjustment action, the photovoltaic inverter power limiting action, and the demand-side response participation action, and the corresponding action is selected according to the current state. Design a composite reward function that integrates scheduling revenue rewards and energy storage lifetime penalties. The scheduling revenue rewards consist of peak-valley arbitrage revenue, ancillary service subsidy revenue, and demand-side response subsidy revenue. The energy storage lifetime penalties consist of the equivalent cycle loss rate and the capacity decay rate. The reinforcement learning scheduling engine is trained offline using a deep deterministic policy gradient algorithm, and an online fine-tuning strategy is used during real-time scheduling to achieve real-time adaptation to battery state of charge, photovoltaic power fluctuations, and grid peak and valley electricity prices.

7. The method according to claim 1, characterized in that, The multi-objective optimization scheduling model also includes a cross-energy collaborative scheduling module, used to access wind power prediction data and gas turbine operating status data to construct a photovoltaic-storage multi-energy collaborative scheduling framework, including: Obtain short-term power forecast data for wind farms, as well as the start-up and shutdown status and upper and lower output limits of gas turbines; To add the sub-objectives of maximizing wind power consumption and minimizing gas turbine operating costs to the objective function of the multi-objective optimization scheduling model; Based on the complementary characteristics of photovoltaic, wind power and gas turbines, the output ratio of each energy source is dynamically allocated. When photovoltaic output is insufficient, gas turbines are used as backup power sources. When photovoltaic output is excessive, energy storage systems are used to absorb the excess power.

8. The method according to claim 1, characterized in that, The step of sending the optimal scheduling strategy to the energy storage converter and photovoltaic inverter for execution, and correcting the optimal scheduling strategy in real time based on the demand-side response signal, includes: Receive demand-side response call signals issued by the power grid dispatching platform, and parse the regulation capacity demand, response time period and incentive price contained in the signals; Determine whether the current energy storage charge state meets the charge level conditions for demand-side response participation. If it does, generate a scheduling correction instruction to participate in demand-side response and superimpose the scheduling correction instruction onto the optimal scheduling strategy. The actual state of charge of the energy storage system is updated based on the demand-side response execution results, and the updated state of charge is used as the state input for the next scheduling cycle, thereby realizing closed-loop interaction with the power grid scheduling platform.

9. The method according to claim 5, characterized in that, The process of solving the multi-objective optimization scheduling model using a non-dominated sorting genetic algorithm based on Pareto optimal solutions includes: Initialize a population with multiple scheduling schemes; iteratively optimize the population through selection, crossover, and mutation operations, and output a Pareto optimal scheduling solution set after multiple generations of evolution; The scheduling decision system selects the optimal scheduling scheme that balances scheduling benefits and energy storage lifetime from the Pareto optimal scheduling solution set based on actual operational needs.

10. A multi-scenario adaptive optical-storage collaborative scheduling system, characterized in that, It includes a data acquisition module, a scene recognition module, a parameter configuration module, a model building module, a reinforcement learning solution module, and an instruction execution module; among which: The data acquisition module is used to acquire photovoltaic power generation prediction data, energy storage system status parameters, and power grid operation status data. The scene recognition module is used to classify and identify the current running scene according to the preset scene recognition model, and determine the current scheduling scene type; The parameter configuration module is used to extract the corresponding scenario scheduling parameter set from a pre-built scenario feature library according to the current scheduling scenario type. The model building module is used to build a multi-objective optimization scheduling model for the current scenario based on the scenario scheduling parameter set. The reinforcement learning solution module is used to input the photovoltaic power generation prediction data, energy storage system state parameters, power grid operation status data and the multi-objective optimization scheduling model into the reinforcement learning scheduling engine for iterative solution and output the optimal scheduling strategy. The instruction execution module is used to send the optimal scheduling strategy to the energy storage converter and photovoltaic inverter for execution, and to modify the optimal scheduling strategy in real time according to the demand-side response signal to generate the final scheduling instruction.

Images