A method for suppressing power fluctuation of electro-thermal hydrogen micro-grid based on deep reinforcement learning

Abstract

The application discloses a power fluctuation suppression method for electrothermal hydrogen micro-grid based on deep reinforcement learning, and belongs to the field of micro-grid energy management. λ L and λ M The fluctuation power is decomposed into low, medium and high frequency components. The low frequency component is distributed to the hydrogen energy storage system in combination with the reference power, the medium frequency component is distributed to the heat pump, and the high frequency component and the unresponsive power in the previous stage are bottomed by the battery. The output of the heat pump and the hydrogen energy storage is respectively linked to the state of the heat storage tank and the battery to be modified. The system has an asymmetric punishment and dynamic auxiliary reward mechanism for the battery: when the battery capacity is under pressure, the intelligent agent is prompted to actively reduce λ L The fluctuation is forced to be transferred to the hydrogen energy storage to protect the battery life. The method releases the auxiliary adjustment potential of multi-energy assets across the frequency band by unbinding the strict constraints of hydrogen heat equipment.

Description

Technical Field

[0001] This invention belongs to the field of microgrid energy management and control technology, and specifically relates to a method for power fluctuation mitigation in electrothermal hydrogen microgrids based on deep reinforcement learning. Background Technology

[0002] With the global energy structure transitioning towards cleaner and lower-carbon energy and the rapid development of distributed generation technologies, the penetration rate of renewable energy sources, such as wind and solar power, in microgrids is increasing. However, the output of these energy sources is characterized by significant randomness, intermittency, and volatility. The large and rapid changes in power pose serious challenges to the power quality (such as voltage and frequency fluctuations), power balance, and harmonious interaction with the main grid. The combined power-thermal-hydrogen microgrid, by integrating the power grid, thermal network, and hydrogen energy network, leverages the flexible conversion, storage, and complementarity between multiple energy sources to provide a key technological approach for improving regional energy utilization efficiency, enhancing system flexibility, and promoting the efficient absorption of renewable energy.

[0003] In the operation and control of a combined electric-thermal-hydrogen microgrid, effectively mitigating power fluctuations introduced by renewable energy generation is one of the core tasks to ensure the stable, reliable, and economical operation of the system. Battery energy storage systems (BESS) have become a common means of mitigating power fluctuations due to their rapid response and bidirectional regulation capabilities. However, if they are left to handle power fluctuations across the entire frequency band, especially at high frequencies and high rates, their cycle life is easily reduced rapidly.

[0004] To address this, multi-energy collaborative power fluctuation mitigation schemes have emerged. These typically combine slow-response but high-capacity hydrogen storage systems (HSS) to handle low-frequency fluctuations, heat pumps (HP) and their associated thermal storage tanks (HST) to handle medium- and low-frequency fluctuations, and battery energy storage systems (BESS) to handle high-frequency fluctuations. Existing technology discloses a strategy for collaboratively mitigating power fluctuations using these multiple types of devices. This strategy is based on frequency division control, employing a low-pass filter with fixed parameters or dynamically adjusted based on simple rules to decompose and distribute the fluctuating power to be mitigated to each device. Simultaneously, through preset logical rules, the output of each unit is constrained and allocated based on the state of charge (SOC) of each device and user-defined temperature dead zones.

[0005] The technical problem to be solved is: how to break away from the dependence on traditional manual experience rules in the face of high-dimensional, nonlinear electrothermal hydrogen coupled microgrid environment, realize data-driven adaptive partitioning of frequency domain intervals, and completely break the rigid state constraints of traditional energy storage equipment to fully release the regulation potential of hydrogen energy and thermal energy assets, so as to protect the battery life to the extreme.

[0006] Although the multi-energy cooperative power mitigation strategies proposed in existing technologies have theoretically verified the feasibility of multiple devices participating in power mitigation, they still face the following challenges in practical engineering applications:

[0007] 1. Poor adaptability of frequency domain partitioning and rigid control strategies: When decomposing fluctuating power, existing technologies rely heavily on manually set fixed parameters or simple fuzzy rules for adjusting the filtering time constant. When dealing with complex electrothermal-hydrogen microgrids with multivariate coupling and varying weather / load conditions throughout the four seasons, traditional methods struggle to find the optimal frequency domain partitioning boundary through high-dimensional state characteristics, thus failing to achieve dynamic and precise power allocation from a global perspective.

[0008] 2. The regulation potential of hydrogen thermal equipment is severely limited: In existing technologies, the power regulation of heat pumps and hydrogen energy storage systems is often directly and strictly limited by the rigid soft / hard constraints of the thermal storage tank / hydrogen energy storage SOC. When the risk of deep charge / discharge of the battery (such as extremely low SOC requiring emergency rescue) conflicts with the demand for microgrid smoothing, traditional control systems and penalty mechanisms often operate independently, hesitant to cross-frequency bands to dispatch thermal and hydrogen equipment for cross-band support. This effectively limits the depth and reliability of hydrogen thermal assets as a flexible resource in emergency response under extreme system conditions. Summary of the Invention

[0009] Purpose of the Invention: This invention aims to overcome the shortcomings of existing multi-energy collaborative power fluctuation mitigation strategies in microgrids, such as the difficulty of handling high-dimensional nonlinear coupling with rule-based control based on human experience, and the poor adaptability of fixed (or simple fuzzy rule) filtering methods under various operating conditions. To this end, this invention proposes a power fluctuation mitigation method for electrothermal hydrogen microgrids based on a Soft Actor-Critic (SAC) deep reinforcement learning algorithm. This method uses the real-time operating status of the system's multi-source heterogeneous systems (including renewable energy fluctuations, real-time state of charge of batteries / hydrogen storage / thermal storage tanks, various time-varying loads, and ambient temperature, etc.) as high-dimensional observation vector inputs. It directly utilizes a neural network to autonomously learn and extract the optimal broadband fluctuation mitigation strategy from massive amounts of operating data with seasonal random meteorological characteristics, completely eliminating the reliance on traditional manual prior logic.

[0010] Technical solution:

[0011] To achieve the above objectives, this invention provides a method for suppressing power fluctuations in an electrothermal hydrogen microgrid based on SAC data-driven frequency domain partitioning, comprising the following steps:

[0012] (1) Comprehensive environmental status observation: The microgrid operating status vector is obtained according to a preset control cycle. The normalized observation parameters include: fluctuation power, SOC. BESS SOC HSS SOC HST, room temperature deviation, hydrogen load, electrical load and outdoor temperature.

[0013] State Acquisition: At discrete time step t, the comprehensive environmental state observation vector of the combined electrothermal-hydrogen microgrid is acquired in real time. The observation vector includes the renewable energy fluctuation power P. Flu Battery State of Charge (SOC) BESS State of Charge (SOC) of hydrogen energy storage systems HSS State of Charge (SOC) of the thermal storage tank HST Indoor temperature deviation, hydrogen load, electrical load, and outdoor temperature.

[0014] (2) SAC dynamically generates filter parameters: The state observation vector is fed into the Actor network of the SAC agent. The network outputs a continuous action vector [-1,1], which is then converted into the first-stage low-pass filter time constant λ through a linear mapping. L ∈[λ L_min , λ L_max ] and the second-stage filter time constant λ M ∈[λ M_min , λ M_max ].

[0015] SAC Action Mapping and Dynamic Decomposition: The observed vectors are input into the Actor network of the trained Flexible Action Evaluation (SAC) algorithm, which outputs continuous action vectors. These action vectors are then mapped to two dynamically adjusted low-pass filter time constants λ. L and λ M A dynamically adjusted low-pass filter is used to suppress the total fluctuating power P. Flu Decomposition, utilizing the first-stage low-pass filter and the time constant λ L Decompose the low-frequency component P L and the remaining component P MH Using a second-stage low-pass filter and the time constant λ M The remaining component P MH Further decomposed into mid-frequency component P M and high-frequency component P H The discretization update formulas for the first-stage and second-stage low-pass filters are as follows: P L (t)=(1-α L )P L (t-1)+ α L P Flu (t),

[0016] P M (t)=(1-α M )P M (t-1)+α M P MH (t); Filtering coefficient αL =Δt / (λ L +Δt), α M =Δt / (λ M +Δt), where Δt is the control step size.

[0017] (3) Third-order frequency domain dynamic decomposition: Based on dynamic λ L and λ M A two-stage first-order low-pass filter is used to filter the net ripple power P. Flu Decompose: P L (Low frequency band): by λ L Extract. P M (Mid-frequency band): The remaining power P to be smoothed MH via λ M Extract. P H (High-frequency band): Total fluctuation minus P L and P M The transient high-frequency margin after.

[0018] Initial allocation of multi-functional devices: The low-frequency component P... L Combine the reference power allocation to the hydrogen energy storage system HSS; and allocate the intermediate frequency component P... M The low-frequency overflow power that the hydrogen energy storage system cannot handle is allocated to the heat pump HP; the high-frequency component P... H All the intermediate frequency overflow power that the heat pump cannot handle is allocated to the battery BESS.

[0019] Coordinated control of hydrogen energy storage system: Obtain the target reference power of the hydrogen energy storage system and linearly amplify it during the soft start phase of initial operation; subsequently, based on the real-time battery state of charge (SOC)... BESS Calculate the derating factor, perform linear interpolation to drate the target reference power, and obtain the final reference power; subtract the final reference power from the original fluctuation power of the microgrid to obtain the total fluctuation power P to be mitigated. Flu Send it to the dynamic decomposition step.

[0020] (4) Power allocation and cascading correction across devices:

[0021] HSS Output Layer: Based on soft-start logic and current SOC BESS Calculate the adaptive reference power. The hydrogen energy storage system assumes the reference power and P. L The sum. If the HSS reaches the physical limit of inflation and deflation, the overflow power P L_nhandled Pass it on to the next level.

[0022] HP output layer: The heat pump command is generated by P that meets the basic load of indoor temperature control. HP_base With the allocated intermediate frequency damping power P M Composed of (including overflow from the previous level). Combined with the SOC of the thermal storage tank. HSTPerform linear extrapolation correction. Overflow power P M_unhandled Passed down to the last level.

[0023] BESS Output Layer: The battery serves as the final safety net, absorbing high-frequency impacts. H and all preamp overflow power.

[0024] (5) SAC Training and Update Based on Customized Reward Function: After the agent interacts with the environment, it enters the simplified hierarchical reward function calculation module. The reward value is composed of the following weighted summation: Fluctuation Smoothing and Temperature Control Reward: Measures the residual fluctuation penalty after microgrid smoothing and the dead zone penalty for deviating from the target room temperature.

[0025] BESS Asymmetric SOC Penalty: To protect the battery, an asymmetric function is used: when SOC... BESS When the value is less than 0.5, a high penalty of a cubic power is imposed to force the agent to recharge; when the value is greater than 0.5, only a mild constraint of a quadratic power is applied.

[0026] The survival red line of complete debinding: only severe penalties are imposed on batteries that touch the 0.1 or 0.95 limit. The SOC anchoring penalty for HSS and HST is completely eliminated, "debinding" hydrogen / thermal equipment from the underlying logic, giving intelligent agents the freedom to freely utilize thermal / hydrogen assets.

[0027] HSS Dynamic Assisted Reward Shaping: When the system senses that the battery is under pressure (such as SOC), BESS When λ < 0.5, the λ output of the actively monitored agent is... L If the agent outputs a smaller λ L (That is, actively allowing the HSS to bear more volatility) will provide a great positive incentive:

[0028]

[0029] Conversely, if the agent allows the HSS to slack off when the battery is low, additional penalties will be imposed.

[0030] Model Training and High-Frequency Battery Protection Mechanism: The training process of the SAC agent employs a customized reward function, which includes fluctuation smoothing rewards, temperature control rewards, battery SOC asymmetric penalty rewards, and survival redline penalties. Among these, the battery SOC asymmetric penalty rewards employ a more severe penalty mechanism than overcharging for battery capacity reduction: when SOC... BESS A cubic exponential penalty is applied when SOC is less than 0.5, and when SOC is less than 0.5... BESS A quadratic mild penalty is applied when the SOC is ≥0.5; a dynamic auxiliary mechanism is also provided: when SOC... BESS When the value is below the safety threshold, the time constant λ is actively reduced for the SAC agent. LPositive rewards are given for the behavior, and the reward guidance system increases the proportion of low-frequency components allocated to the hydrogen energy storage system in order to forcibly transfer fluctuations and actively share the pressure on the battery; Command issuance: The final system coordinated control command is issued to the corresponding hydrogen energy storage system, heat pump and battery for execution.

[0031] Beneficial effects:

[0032] Compared with existing technologies, this invention achieves the following significant technological advancements and beneficial effects through a unique electrothermal hydrogen multi-energy microgrid system configuration (HSS+HP+HST+BESS) and an innovative data-driven collaborative control strategy (dynamic frequency domain partitioning based on SAC algorithm, battery asymmetric penalty, and dynamic hydrogen thermal assistance):

[0033] 1. Breaking through rule limitations to achieve adaptive and precise frequency domain partitioning: This invention abandons the traditional fixed filtering or simple fuzzy rules that rely on manual experience, and utilizes deep reinforcement learning (SAC) algorithms to directly learn from the massive microgrid operating status features. The system can dynamically and continuously output the time constant of a two-stage low-pass filter based on multi-dimensional heterogeneous real-time operating conditions (such as meteorological randomness, real-time SOC of multi-source equipment, etc.), thereby accurately partitioning the total fluctuating power into three-order frequency domain components ("low, medium, and high"), greatly improving the system's adaptive capability and scientific allocation in complex nonlinear coupling environments.

[0034] 2. Fully unleash the regulatory potential of hydrogen thermal assets and achieve deep synergy across the entire frequency band: This invention creatively removes the rigid dead-zone anchoring penalty imposed by traditional methods on the state of charge of hydrogen storage (HSS) and thermal storage tanks (HST), limiting them only to physical baseline constraints. This grants the deep reinforcement learning agent a high degree of freedom in action exploration, enabling the slower-responding hydrogen energy and thermally inertial heat pumps not only to handle low-to-medium frequency fluctuations but also to be fully utilized in the global energy balance of the microgrid under extreme operating conditions, maximizing the utilization rate of multi-energy assets.

[0035] 3. Constructing a proactive defense mechanism to achieve maximum battery life extension and system cost reduction: This invention targets the battery energy storage system (BESS), a core and vulnerable asset, and innovatively develops a customized multi-objective reward function. On one hand, it strictly curbs deep battery discharge through SOC asymmetric penalties; on the other hand, it pioneers a Reward Shaping dynamic assistance mechanism. When battery pressure is detected, the system can guide the agent to actively break through conventional frequency domain boundaries through positive rewards, driving hydrogen and thermal energy devices to share the high-frequency impact. This mechanism, without requiring additional expensive high-frequency energy storage hardware such as supercapacitors, constructs a cross-device proactive rescue system for batteries, significantly delaying battery cycle life degradation and improving the overall long-term operational reliability and economic benefits of the microgrid. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. It should be understood that the drawings are only schematic and are not drawn to scale.

[0037] Figure 1 This is a schematic diagram of the logic for third-order frequency domain dynamic decomposition and multi-energy device coordinated allocation based on SAC intelligent agents provided by the present invention. Detailed Implementation

[0038] This invention provides a method for power fluctuation mitigation in an electrothermal hydrogen microgrid based on deep reinforcement learning. This method is based on a Soft Actor-Critic (SAC) intelligent agent model, which is periodically executed within the microgrid's operating cycle. Specifically, it includes the following steps: (1) Comprehensive environmental state parameter acquisition step: At discrete time step t, the controller monitors in real time and acquires the current operating state parameters of the electrothermal hydrogen joint microgrid system with a preset control step size, constructing a comprehensive environmental state observation vector. The state observation vector output by the microgrid environment, after normalization, specifically includes: the total fluctuation power P of renewable energy to be mitigated at the current moment. Flu Battery current state of charge (SOC) BESS Current State of Charge (SOC) of Hydrogen Energy Storage System HSS Current State of Charge (SOC) of the thermal storage tank HST The indoor temperature deviation from the standard room temperature, the predicted hydrogen load, the predicted conventional electrical load, and the current outdoor ambient temperature T out .

[0039] (2) Dynamic filtering and fluctuation decomposition steps based on SAC agent: combining Figure 1 The third-order frequency domain dynamic decomposition logic shown inputs the state observation vector acquired in step (1) into the Actor network of the trained SAC agent. The network outputs a continuous action vector in the range [-1, 1], which the system converts into the dynamic time constant λ of the first-stage low-pass filter (LPF1) through action mapping. L (Mapped to [λ]) L_min , λ L_max The dynamic time constant λ of the second-stage low-pass filter (LPF2) and the second-stage low-pass filter. M (Mapped to [λ]) M_min , λ M_max The dynamic time constant is used to adjust the total fluctuation power P to be smoothed in step (1). Flu Perform third-order frequency domain decomposition. Calculate the low-frequency ripple component P after the first-stage low-pass filter. L (Low Freq), and the remaining component P to be mitigated. MHSubsequently, P MH The intermediate frequency fluctuation component P is extracted again through a second-stage low-pass filter. M (MidFreq), and the high-frequency fluctuation component P was extracted. H (High Freq) is P H = P MH - P M The discretized filtering formulas for the first and second stage low-pass filters are as follows: P L (t)=(1-α L )P L (t-1)+ α L P Flu (t), P M (t)=(1-α M )P M (t-1)+ α M P MH (t); where α L With α M Using λ respectively L and λ M The dynamic filtering coefficient α is obtained by combining the control step size Δt. L =Δt / (λ L +Δt), α M =Δt / (λ M +Δt).

[0040] (3) Preliminary allocation and coordination steps for the hydrogen energy storage system (HSS): Calculate the baseline power that the hydrogen energy storage system should handle. This step introduces a soft-start mechanism and adaptive derating logic based on battery state: The baseline power is linearly amplified for a soft start within the first 60 minutes of system operation; simultaneously, based on the current SOC... BESS Calculate the depreciation factor (when SOC) BESS When the depreciation factor is ≥0.6, the depreciation factor is 1. BESS When the value is ≤0.4, it is 0, and the value in between is a linear interpolation). The power after soft start is multiplied by this derating factor to obtain the final reference power. The original fluctuation power of the microgrid is subtracted from the final reference power, which is taken as the total fluctuation power P to be smoothed in step (2) above. Flu Send it to the dynamic decomposition step. For example... Figure 1 As shown, the hydrogen energy storage system preferentially handles the reference power and low-frequency fluctuation component P. L The sum of these parameters, constrained by the maximum charge / discharge power and physical capacity (0.05 to 0.95), is used to calculate the actual output hydrogen storage power P. HSS The low-frequency overflow power that the hydrogen energy storage system cannot handle due to physical constraints, i.e., the unresponsive power, is denoted as P. L_unhandled It is then passed to the heat pump unit for processing.

[0041] (4) Heat pump (HP) output coordination control steps: Perform refined calculations of the heat pump's power command. First, based on the indoor temperature T... in Determine if the temperature is within the comfort zone. If it is, set the target temperature T. tgt Set the current room temperature to maintain indoor heat balance; otherwise, anchor the target temperature at the corresponding upper and lower limits. Calculate the total indoor heat demand to maintain this target through feedforward (to offset wall heat loss) and feedback (room temperature difference), and then convert it into the heat pump's base operating power command P. HP_base Based on this, such as Figure 1 As shown, the heat pump needs to bear the allocated mid-frequency fluctuation component P. M P transferred from the previous level L_unhandled The sum. Combined with the current SOC (State of Charge) of the thermal storage tank. HST The correction coefficient α is calculated by linear extrapolation interpolation to correct the fluctuation smoothing part, thus obtaining the actual power P of the heat pump used to smooth the fluctuation. HP_flu The final master command issued to the heat pump is P. HP_cmd = P HP_base + P HP_flu The unresponsive power, which the heat pump cannot handle due to its constraints, is denoted as P. M_unhandled It is then transmitted to the battery unit.

[0042] (5) Battery Escalator (BESS) Output Backup and Command Issuance Steps: (e.g.) Figure 1 As shown, the battery, as the final backup device, bears the high-frequency fluctuation component P. H And the unresponsive power P transferred from the heat pump M_unhandled After considering the free-running dead zone, a preliminary total power command for the battery is generated. Based on its maximum charge / discharge power limit and strict SOC constraints, the final command P issued to the battery converter is calculated. BESS .

[0043] (6) Model Training and High-Frequency Battery Protection Mechanism Steps: During the offline training and online fine-tuning phases of the model, the system extracts data at the end of the control cycle, extracting system physical response data such as the output of each device, and performs the calculation of a customized multi-objective reward function. The reward / penalty for each sub-item is weighted and summed to obtain the final reward value R. This reward value R is fed back to the Critic network of the SAC agent for model updates, guiding the Actor network to continuously optimize. The specific sub-item calculation rules of the customized reward function are as follows:

[0044] (a) Fluctuation Smoothing Penalty: The penalty is based on the square of the residual between the final output of each device and the total fluctuation. The specific formula is: -(residual fluctuation / 0.5)².

[0045] (b) Temperature control penalty: Anchored at 23℃, with a 0.2℃ no-penalty dead zone. When the deviation is greater than 0.2℃, the specific formula is: -(ABS(T in - 23)- 0.2)².

[0046] (c) BESS Asymmetric SOC Penalty (Core): Implements a more stringent protection mechanism than overcharging for battery capacity degradation. (SOC determination) BESS Does < 0.5 hold? If so, it indicates the system is in a deep power-feeding danger, and a cubic exponential penalty is imposed, specifically: -((0.5- SOC) BESS If the value is 0.1), force the intelligent agent to charge quickly or reduce its output; otherwise (i.e., SOC). BESS If the value is ≥0.5, then only a light quadratic penalty is applied, specifically: -((SOC) BESS - 0.6) / 0.2)².

[0047] (d) HSS Dynamic Assisted Reward (Reward Shaping Core): When determining SOC BESS When the value is below a safety threshold (e.g., 0.5), evaluate λ. L Normalized value. If the agent's output λ L If the load is relatively small (i.e., the HSS actively shares more of the load fluctuations to assist the battery), then a positive incentive (r) is given. assist > 0); if the agent outputs λ L If the value is large (i.e., without auxiliary battery), an additional penalty (r) will be applied. assist < 0). This mechanism guides the system to actively share the pressure on the battery.

[0048] (e) Survival Red Line Penalty: Severe penalties are imposed only on devices that reach physical limits. Heavy penalties are applied when BESS SOC ≤ 0.1 or BESS SOC ≥ 0.95; simultaneously, the strict electrical anchoring penalties for HSS and HST are lifted, limiting them only to physical overflow prevention boundaries, i.e., physical overflow prevention penalties are applied when HSS / HST SOC ≤ 0.05 or HSS / HST SOC ≥ 0.95. This rule effectively unlocks the cross-band auxiliary regulation potential of multi-energy assets.

Claims

1. A method for mitigating power fluctuations in an electrothermal hydrogen microgrid based on deep reinforcement learning, characterized in that, The method includes state acquisition: at discrete time step t, a comprehensive environmental state observation vector of the combined electrothermal-hydrogen microgrid is acquired in real time, the observation vector including the renewable energy fluctuation power P. Flu Battery State of Charge (SOC) BESS State of Charge (SOC) of hydrogen energy storage systems HSS State of Charge (SOC) of thermal storage tank HST Indoor temperature deviation, hydrogen load, electrical load, and outdoor temperature.

2. The method according to claim 1, characterized in that, This includes SAC action mapping and dynamic decomposition: the observed vectors are input into the Actor network of the trained Flexible Action Evaluation (SAC) algorithm, which outputs continuous action vectors and maps these action vectors to two dynamically adjusted low-pass filter time constants λ. L and λ M A dynamically adjusted low-pass filter is used to suppress the total fluctuating power P. Flu Decomposition, utilizing the first-stage low-pass filter and the time constant λ L Decompose the low-frequency component P L and the remaining component P MH Using a second-stage low-pass filter and the time constant λ M The remaining component P MH Further decomposed into mid-frequency component P M and high-frequency component P H The discretization update formulas for the first-stage and second-stage low-pass filters are as follows: P L (t)=(1-α L )P L (t-1)+ α L P Flu (t), P M (t)=(1-α M )P M (t-1)+α M P MH (t); Filtering coefficient α L =Δt / (λ L +Δt), α M =Δt / (λ M +Δt), where Δt is the control step size.

3. The method according to claim 2, characterized in that, This includes the initial allocation of multi-functional devices: including the low-frequency component P. L Combine the reference power allocation to the hydrogen energy storage system HSS; and allocate the intermediate frequency component P... M The low-frequency overflow power that the hydrogen energy storage system cannot handle is allocated to the heat pump HP; the high-frequency component P... H All the intermediate frequency overflow power that the heat pump cannot handle is allocated to the battery BESS.

4. The method according to claim 3, characterized in that, This includes coordinated control of the hydrogen energy storage system: obtaining the target reference power of the hydrogen energy storage system and linearly amplifying it during the soft-start phase of initial operation; subsequently, based on the real-time state of charge (SOC) of the battery... BESS Calculate the derating factor, perform linear interpolation to drate the target reference power, and obtain the final reference power; subtract the final reference power from the original fluctuation power of the microgrid to obtain the total fluctuation power P to be mitigated. Flu Send it to the dynamic decomposition step.

5. The method according to claim 4, characterized in that, Including model training and high-frequency battery protection mechanisms: The training process of the SAC agent adopts a customized reward function, which includes fluctuation smoothing rewards, temperature control rewards, battery SOC asymmetric penalty rewards, and survival redline penalties; among them, the battery SOC asymmetric penalty rewards set a more severe penalty mechanism than overcharging for battery power decline: when SOC... BESS A cubic exponential penalty is applied when SOC is less than 0.5, and when SOC is less than 0.5... BESS A quadratic mild penalty is applied when the SOC is ≥0.5; a dynamic auxiliary mechanism is also provided: when SOC... BESS When the value is below the safety threshold, the time constant λ is actively reduced for the SAC agent. L Positive rewards are given for the behavior, and the reward guidance system increases the proportion of low-frequency components allocated to the hydrogen energy storage system in order to forcibly transfer fluctuations and actively share the pressure on the battery; Command issuance: The final system coordinated control command is issued to the corresponding hydrogen energy storage system, heat pump and battery for execution.

Images