Energy management method for solid oxide fuel cell microgrid

Abstract

The application discloses an energy management method suitable for a solid oxide fuel cell microgrid, and belongs to the field of fuel cell new energy sources. The application comprises the following steps: establishing a microgrid energy management optimal control problem considering two optimization indexes of the economy of the microgrid and the durability of the SOFC; and solving the energy management optimal control problem by splitting it into two time scales of day-ahead planning with days as the cycle and real-time scheduling with minutes as the unit. In the day-ahead planning stage, the day-ahead net power data predicted based on the long short-term memory neural network is used to plan the SOFC power generation in the next 24 hours by using the deep deterministic policy gradient reinforcement learning algorithm, so as to provide the optimal initial value for the real-time scheduling and reduce the calculation requirement of the real-time optimization; and in the real-time scheduling stage, the generalized model predictive control algorithm is used for real-time scheduling, so as to eliminate the prediction error in the day-ahead planning stage and balance the power demand in real time.

Description

Technical Field

[0001] This invention belongs to the field of fuel cell new energy technology, specifically relating to an energy management method suitable for microgrids based on solid oxide fuel cells. Background Technology

[0002] A microgrid is a small-scale power generation and distribution system composed of distributed power sources, energy storage devices, energy conversion devices, loads, monitoring and protection devices, etc. To overcome the intermittency and randomness of renewable energy (such as photovoltaic and wind turbine) power generation due to weather conditions, energy storage devices (such as lithium batteries and electrolyzers), auxiliary power generation systems (such as solid oxide fuel cells) are typically combined with renewable energy sources to form a microgrid system to meet time-varying load demands. While meeting load demands, various power allocation schemes exist due to the coupling and interaction between multiple energy devices, and different power allocation schemes have different operating costs, efficiency, and other operating indicators. Therefore, energy management strategies are needed to appropriately coordinate the power allocation ratio among various energy devices to achieve the optimal operating cost and efficiency.

[0003] Currently, microgrid energy management strategies are mainly divided into three categories: rule-based, optimization-based, and learning-based. Rule-based energy management strategies rely on power allocation rule tables designed with expert experience, making it difficult to achieve optimal results for different load demand scenarios. Optimization-based energy management strategies require highly accurate microgrid dynamic models; as the number of microgrid devices increases, the modeling difficulty and computational burden also increase, making it difficult to obtain the optimal power allocation for the above objectives in real time while satisfying complex operational constraints. Learning-based energy management strategies rely on actual operating data to adjust their algorithm parameters and decision-making processes to achieve the optimal results. For microgrids with complex nonlinear solid oxide fuel cells, although the above energy management strategies can provide power allocation that meets load demand, it is difficult to obtain the power allocation with optimal operating cost and efficiency in real time under the influence of the intermittency and randomness of renewable energy generation. Summary of the Invention

[0004] The purpose of this invention is to provide an energy management method suitable for microgrids based on solid oxide fuel cells, in order to obtain power allocation with optimal operating cost and efficiency in real time, addressing the problems mentioned above.

[0005] To achieve the above-mentioned objectives, the present invention provides a multi-timescale energy management method for microgrids based on solid oxide fuel cells, comprising two stages: day-ahead planning and real-time scheduling. In the day-ahead planning stage, power planning is performed based on the net power data predicted for the day, and the optimal power reference point for the next 24 hours is used as the reference value for real-time power allocation to reduce the computational requirements for real-time optimization. In the real-time scheduling stage, the influence of prediction errors is eliminated to ensure the balance of power allocation in the microgrid.

[0006] The energy management method for solid oxide fuel cell microgrids provided by this invention includes the following steps:

[0007] Step 1: Establish the optimal control objective function for microgrid energy management (i.e., the optimal control problem). The optimization indices of this objective function include: economic indicators and durability indicators.

[0008] The constraint condition for optimizing the objective function is: the photovoltaic power P of the photovoltaic cell. pv The wind turbine power P of a wind turbine generator wt The net power P of a solid oxide fuel cell (SOFC) system fc Lithium battery power P bat Hydrogen production equipment power P el The sum of the load power demand P load Phase equilibrium;

[0009] Furthermore, the microgrid is an isolated DC microgrid system with photovoltaic cells, small wind turbine generators, solid oxide fuel cells, lithium batteries, and electrolyzers;

[0010] Step 2: Decompose the solution of the established optimization objective function into two time scales: daily planning with a period of one day and real-time scheduling with a period of one minute.

[0011] Step 3: For the day-ahead planning timescale of the optimization objective function, use a long short-term memory neural network to calculate the day-ahead power generation P of renewable energy. PV P wt and load power demand P load Perform day-ahead power forecasting;

[0012] Step 4: For the day-ahead planning timescale of the optimization objective function, based on the load power demand P load The predicted power of the SOFC system in the next 24 hours is obtained by using a deep deterministic policy gradient reinforcement learning algorithm. fc A day-ahead forecast was performed to obtain the net power P. fc The current forecast results;

[0013] Step 5: For the real-time scheduling timescale of the optimization objective function, apply the generalized model predictive control algorithm to the net power P. fc The system performs real-time scheduling based on the day-ahead forecast results, eliminating errors in the forecast results at each day-ahead scheduling moment to balance real-time load power demand.

[0014] Furthermore, the specific objective function for optimal control of microgrid energy management is set as follows:

[0015]

[0016] sty=P pv +P wt +P fc +P bat +P el =P load

[0017] Where u is the first control action in the optimal control strategy set, t0 is the initial time, and t is the current time; T EV =N×Δt is the entire scheduling period, Δt is the sampling interval, and N is the number of sampling points included in the scheduling period; a1 and a2 are J econ (t) and J life The preset weighting coefficients of (t) are given, and the sum of a1 and a2 is 1. By adjusting the ratio of the weighting coefficients, the priority of economic indicators and durability indicators can be adjusted. P pv Photovoltaic power, P wt P represents the power of the wind turbine. fc P represents the net power of the SOFC system. bat For lithium battery power, P el For the power of hydrogen production equipment, P load This refers to the power requirements of the load.

[0018] For the safety of microgrid operation, the following constraints are considered:

[0019] (1) Power range limitation: |P bat |≤N bat P r,bat 0≤P fc ≤N fc P r,fc -N el P r,el ≤P el ≤0;

[0020] (2) Power variation limit: |ΔP bat |≤0.5N bat P r,bat 、|ΔP fc |≤0.25N fc Pr,fc 、|ΔP el |≤0.5N el P r,el ;

[0021] (3) Storage capacity limit: 0.2≤SoC≤0.8, 0.05≤LoH≤0.95.

[0022] Where, N i and P r,i Let pv, wt, fc, bat, el, tank be the number of devices and the rated output power of the i-th object, respectively, where i ∈ {pv, wt, fc, bat, el, tank}, pv, wt, fc, bat, el, tank are photovoltaic, wind turbine, fuel cell, lithium battery, electrolyzer and hydrogen storage tank, respectively, SoC is the battery state of charge and LoH is the hydrogen level.

[0023] Furthermore, economic indicators J econ (t) and durability index J life (t) Specifically:

[0024] J life =C inv,fc (N fc V d ;

[0025] Where t represents time, LHV H2 Because of the low calorific value of hydrogen, η fc (·) is the SOFC efficiency function, V d This represents the SOFC voltage degradation rate.

[0026] Furthermore, in step 2, by setting the time parameter in step 1, the optimal control problem established in step 1 is decomposed into two time scales for solution: day-ahead planning with a period of one day and real-time scheduling with a period of one minute. Specifically:

[0027] (1) Day-ahead planning time scale: The period is 24 hours and the interval is 1 hour, that is, setting T in the objective function. EV =24, Δt=1, N=24. First, a Long Short-Term Memory Neural Network (LSTM) is used to predict the day-ahead net power (step 3). Then, a deep deterministic policy gradient reinforcement learning algorithm is used to plan the day-ahead SOFC power generation for the next 24 hours (step 4).

[0028] (2) Real-time scheduling time scale: The period is 1 hour and the interval is 5 minutes, that is, the T in the objective function is set. EV =1 hour = 60, Δt = 5, N = 12. Real-time scheduling is performed using a generalized model predictive control algorithm to eliminate prediction errors in day-ahead scheduling (step 5).

[0029] Furthermore, in step 3, the daily power generation P of renewable energy is... PV P wt and load power demand P load The specific prediction is as follows:

[0030]

[0031] Where f(·) is the output of the LSTM neural network model, x represents the load power for the next 24 hours. d ={x d,0 ,x d,1 ,..,x d,23} represents the prediction input for day d, x d,t Let x be the eigenvector of the t-th hour on day d, where t = {0, 1, ..., 23}. d,t The characteristic expression considers factors including photovoltaic and wind turbine power output and load demand, as well as historical data on solar radiation, wind speed, temperature, load demand, and future weather types. Its characteristic expression is:

[0032] x d,t ={(d+t / 24) / 365,R d,t ,ν d,t ,P d,t ,T d+1,t ,Td d+1,t B d+1,t H d+1,t W d+1,t C d+1,t ,S d+1,t}

[0033] Where d represents the day of the year, R d,t v d,t and P d,t Given the solar radiation, wind speed, and load requirements for hour t on day d, T d+1,t 、Td d+1,t B d+1,t H d+1,t W d+1,t C d+1,t and S d+1,t These are the forecast data for temperature, dew point temperature, air pressure, humidity, weather type, cloud cover, and visibility for the next t hours.

[0034] Furthermore, in step 4, the state, action, and reward in the deep deterministic policy gradient reinforcement learning algorithm are:

[0035] Action a d ={P fc,0 ,P fc,1 ,…,Pfc,23} is defined as the power generation capacity of SOFC in the next 24 hours;

[0036] The state is defined as: s d ={SoC d,0 LoH d,0 ,P d,0 ,P d,1 ,...,P d,23};

[0037] Among them, P fc,t For the net power in the next t = {0, 1, ..., 23} hours, SoC d,0 and LoH d,0 The initial SoC and LoH values ​​of the microgrid at time 0 on day d; P d,t For the load power requirement P load The day-ahead forecast power value (i.e., the day-ahead forecast net power value) for the d-th hour;

[0038] The reward is defined as:

[0039]

[0040] Furthermore, to ensure the rationality of the actions, a large negative reward is provided and the training segment is terminated when the action violates energy storage and power flow constraints. After obtaining the SOFC power generation plan, the energy of the lithium battery, electrolyzer, and hydrogen storage tank is determined by the following three rules:

[0041] (1) Lithium batteries have higher charging efficiency, so when the net power is greater than zero, lithium batteries are charged first, and hydrogen is electrolyzed after they are fully charged.

[0042] (2) When the sum of the power generated by SOFC and the net power is still less than zero, the power balance is maintained by the discharge of lithium batteries.

[0043] (3) SOFCs should use electrolyzed hydrogen first, and add (purchase) hydrogen only when necessary.

[0044] Furthermore, in step 5, the prediction, correction, and optimization steps of the generalized model predictive control algorithm are as follows:

[0045] (5-1) Power Prediction:

[0046] The sum of photovoltaic and wind turbine power generation at p future moments, P gen The regression model used for prediction is expressed as the following equation:

[0047]

[0048] Among them, among them, P at time k to time k+j genThe predicted value, p is the forward prediction step size, j={0,2,...,p-1}, τ is the past data step size, β0~β τ This represents τ+1 preset coefficients;

[0049] The predicted output power of the microgrid is:

[0050]

[0051] In the formula, P fc (k+j|k), P bat (k+j|k) and P el (k+j|k) represents the control power of the fuel cell, lithium battery, and electrolyzer at time k+j, respectively, which is the corresponding value at time k+j based on time k.

[0052] (5-2) Feedback correction:

[0053] The forecast disturbances caused by the uncertainties in solar radiation and wind speed are The prediction error can then be corrected in the following ways:

[0054]

[0055] Among them, P gen (k) represents the actual power generation at time k, i.e., P gen The current actual value, P represents the predicted power generation at time k. gen The current predicted value (based on time k predicting time k+j), The corrected predicted value; H is a constant between 0 and 1, preferably H = 0.5;

[0056] (5-3) Rolling optimization:

[0057] Using a penalty function, the day-ahead planning reference value P is... fc,ref Add (k+j|k) to the real-time scheduler to construct a scheduling optimization problem:

[0058]

[0059]

[0060] Where k represents the current time, P load (k|k) represents the load power demand at time k, P fc (k+j|k) represents the corresponding value of the net power of SOFC for the next 24 hours obtained through reinforcement learning, i.e., a d ={P fc,0 ,P fc,1 ,...,Pfc,23 The value corresponding to}; based on the scheduling optimization problem, the control quantity is solved to obtain the optimal control power of the fuel cell, lithium battery, and electrolyzer at several times after the current time k, thus obtaining the optimal control strategy set. All the optimal control power at each time after time k is taken as a control action. The control actions in the optimal control strategy set are arranged in descending order of time, and the first control action in the optimal control strategy set is applied to the microgrid. That is, solving for the control quantity (P) fc P bat P el After obtaining the optimal values ​​of the next few time steps from the current time k, the solved control variables constitute the optimal control strategy set, but only the optimal values ​​of the next time step from time k are used as control actions and applied to the microgrid.

[0061] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0062] (1) Based on a multi-timescale energy management strategy, this invention provides the optimal reference point for real-time scheduling by performing day-ahead planning. In addition to reducing the amount of computation required for real-time scheduling, it also performs a global search for the optimal power allocation on a one-day cycle, thus avoiding the impact of the intermittency of renewable energy on the cost optimization of microgrids.

[0063] (2) The present invention takes into account the degradation cost of SOFC in microgrids, further improving the economy of microgrids. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of a microgrid system based on a solid oxide fuel cell constructed according to an embodiment of the present invention;

[0065] Figure 2 This is the multi-timescale energy management proposed in the embodiments of the present invention;

[0066] Figure 3 This is a schematic diagram of the processing results of an embodiment of the present invention, wherein 3(a) is the day-ahead net power and its prediction results in an embodiment of the present invention; 3(b) is the day-ahead scheduling results based on the day-ahead 24-hour net power prediction;

[0067] Figure 4 This is a schematic diagram of the power balance error processing of the day-ahead scheduling result in an embodiment of the present invention, wherein 4(a) is the power balance error of the day-ahead scheduling result in an embodiment of the present invention; 4(b) is the corrected microgrid power balance error in an embodiment of the present invention.

[0068] Figure 5 This is the power allocation result of multi-timescale energy management that combines day-ahead planning with real-time scheduling in the embodiments of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0070] This invention provides an energy management method applicable to microgrids based on solid oxide fuel cells (SOFCs). The method first establishes an optimal control problem for microgrid energy management, considering both the economic efficiency of the microgrid and the durability of the SOFC. This problem includes an isolated DC microgrid system model with photovoltaic cells, small wind turbines, solid oxide fuel cells, lithium batteries, and an electrolyzer. Then, the solution to this optimal control problem is broken down into two time scales: day-ahead planning with a daily cycle and real-time scheduling with a minute-ahead cycle. (1) In the day-ahead planning stage, based on the day-ahead net power data predicted by a long short-term memory neural network, a deep deterministic strategy gradient reinforcement learning algorithm is used to plan the SOFC power generation for the next 24 hours, providing the optimal initial value for real-time scheduling and reducing the computational requirements for real-time optimization. (2) In the real-time scheduling stage, a generalized model predictive control algorithm is used for real-time scheduling to eliminate prediction errors in the day-ahead planning stage and balance power demand in real time.

[0071] To implement the energy management method for solid oxide fuel cell-based microgrids provided in this invention, firstly, a microgrid system model for energy management is constructed, comprising photovoltaic cells, a small wind turbine generator, a solid oxide fuel cell, a lithium battery, and an electrolyzer. This system model can be divided into four subsystems: a renewable energy generation system (including photovoltaic cells and wind turbine generators), a residential power system, an auxiliary power generation system (including fuel cells), and an energy storage system (including lithium batteries, an electrolyzer, and an oxygen storage tank). Figure 1 As shown.

[0072] As one possible implementation, such as Figure 2 As shown, the specific implementation steps of the energy management method for microgrids based on solid oxide fuel cells provided in this embodiment of the invention include:

[0073] S1. Establish the optimal control problem of microgrid energy management considering two optimization indices: the economy of microgrid and the durability of SOFC (i.e., the optimal control objective function of microgrid energy management).

[0074] The expression for the optimal control problem of microgrid energy management is as follows:

[0075]

[0076] sty=P pv +P wt +Pfc +P bat +P el =P load

[0077] Where t0 is the initial time, t is the current time; T EV =N×Δt is the entire scheduling period, Δt is the sampling interval, and N is the number of sampling points included in the scheduling period; a1 and a2 are J econ (t) and J life The preset weighting coefficients of (t) sum to 1. By adjusting the proportion of the weighting coefficients, the priority of economic indicators and durability indicators can be adjusted; P pv Photovoltaic power, P wt P represents the power of the wind turbine. fc P represents the net power of the SOFC system. bat For lithium battery power, P el For the power of hydrogen production equipment, P load This refers to the power requirements of the load.

[0078] For the safety of microgrid operation, the following constraints are considered:

[0079] 1. Power range limitation: |P bat |≤N bat P r,bat 0≤P fc ≤N fc P r,fc -N el P r,el ≤P el ≤0;

[0080] 2. Power variation limit: |ΔP bat |≤0.5N bat P r,bat 、|ΔP fc |≤0.25N fc P r,fc 、|ΔP el |≤0.5N el P r,el ;

[0081] 3. Storage capacity limits: 0.2≤SoC≤0.8, 0.05≤LoH≤0.95.

[0082] Where, N j P r,j j = {pv,wt,fc,bat,el,tank} represents the number of photovoltaic, wind turbine, fuel cell, lithium battery, electrolyzer, and hydrogen storage tank equipment and their rated output power, respectively; SoC represents the battery state of charge, and LoH represents the hydrogen level.

[0083] The economics of SOFC J econ and durability J life Specifically:

[0084]

[0085] Among them, LHV H2 The lower heating value of hydrogen; η fc (·) is the SOFC efficiency function.

[0086] J life =C inv,fc (N fc V d

[0087] Wherein, V d This represents the SOFC voltage degradation rate.

[0088] S2. The above energy management optimal control problem is broken down into two time scales: day-ahead planning with a period of one day and real-time scheduling with a period of one minute. Specifically:

[0089] (1) Day-ahead planning time scale: The period is 24 hours and the interval is 1 hour, that is, setting T in the objective function. EV =24, Δt=1, N=24. First, a long short-term memory neural network is used to predict the day-ahead net power, see step S3. Then, a deep deterministic policy gradient reinforcement learning algorithm is used to plan the day-ahead SOFC power generation for the next 24 hours, see step S4.

[0090] (2) Real-time scheduling time scale: The period is 1 hour and the interval is 5 minutes, that is, the T in the objective function is set. EV =1, Δt=5, N=12. Real-time scheduling is performed using a generalized model predictive control algorithm to eliminate prediction errors in day-ahead scheduling, see step S5.

[0091] S3. Predicting day-ahead net power using a long short-term memory neural network:

[0092]

[0093] Where f(·) is the output of the LSTM neural network model, Net power for the next 24 hours; x d ={x d,1 ,x d,1 ,…,x d,23} represents the prediction input for day d, x d,tt = {0, 1, ..., 23} is the feature vector. Considering that the power output and load demand of photovoltaic and wind turbines are related to historical data of solar radiation, wind speed, temperature, load demand, and future weather types, it is defined in this embodiment as:

[0094] x d,t ={(d+t / 24) / 365,R d,t ,ν d,t ,P d,t ,T d+1,t ,Td d+1,t B d+1,t H d+1,t W d+1,t C d+1,t ,S d+1,t}

[0095] Where d represents the day of the year; R d,t v d,t and P d,t This provides data on solar radiation, wind speed, and load demand for the current day, hour t. d+1,t 、Td d+1,t B d+1,t H d+1,t W d+1,t C d+1,t and S d+1,t These are the forecast data for temperature, dew point temperature, air pressure, humidity, weather type, cloud cover, and visibility for the next t hours. This implementation uses hourly historical meteorological data from Chengdu Shuangliu International Airport, China, from January 1, 2018 to October 30, 2022, as the weather forecast data. Finally, the day-ahead net power and its forecast results for the location on October 25, 2022 are as follows: Figure 3 As shown in 3(a) of the diagram.

[0096] S4. Based on the day-ahead net power forecast, a deep deterministic policy gradient reinforcement learning algorithm is used to plan the SOFC power generation for the next 24 hours. d ={P fc,0 ,P fc,1 ,…,P fc,23} represents the power generation capacity of the SOFC in the next 24 hours; the state is defined as:

[0097] s d ={SoC d,0 LoH d,0 ,P d,0 ,P d,1 ,...,P d,23}

[0098] Among them, SoC d,0and LoH d,0 The initial SoC and LoH values ​​of the microgrid at time 0 on day d; P d,t t = {0, 2, ..., 23} represents the net power value predicted a day ago;

[0099] The reward is defined as:

[0100]

[0101] Furthermore, to ensure the rationality of the actions, a large negative reward is provided and the training segment is terminated when an action violates energy storage and power flow constraints. Finally, the day-ahead scheduling results based on the 24-hour net power prediction are as follows: Figure 3 As shown in Figure 3(b).

[0102] S5. Real-time scheduling is achieved through a generalized model predictive control (GMDC) algorithm to eliminate prediction errors in day-ahead forecasts and balance real-time power demand. The GMDC algorithm for solving real-time scheduling involves three steps: prediction, correction, and optimization, as detailed below:

[0103] (1) Power prediction

[0104] The sum of photovoltaic and wind turbine power generation at p future moments, P gen The regression model used for prediction is expressed as the following equation:

[0105]

[0106] in, Let p be the predicted sum of photovoltaic and wind turbine power generation at time k+j from time k; τ be the forward prediction step size; β0~β τ This represents τ+1 preset coefficients.

[0107] The predicted output power of the microgrid is:

[0108]

[0109] In the formula, P fc (k+j|k), P bat (k+j|k) and P el (k+j|k) represents the control power of the fuel cell, lithium battery, and electrolyzer at time k to time k+j, respectively.

[0110] (2) Feedback correction

[0111] The forecast disturbances caused by the uncertainties in solar radiation and wind speed are The prediction error can then be corrected in the following ways:

[0112]

[0113] In the formula, The corrected predicted value; H is a constant between 0 and 1, and in this example, H = 0.5;

[0114] (3) Rolling optimization

[0115] Using a penalty function, the day-ahead planning reference value P is... fc,ref Adding (k+j|k) to the real-time scheduling process, the optimization problem becomes:

[0116]

[0117]

[0118] After solving for the control variables, only the first control action of the optimal control strategy set is applied to the microgrid.

[0119] in, Figure 4 A schematic diagram of the power balance error processing for the day-ahead scheduling results in this embodiment is provided. Figure 4 4(a) in the figure represents the power balance error between the actual photovoltaic power generation, wind turbine power generation, and load data varying on a minute-by-minute scale and the day-ahead dispatch results on an hourly scale. Figure 4 4(b) in the figure represents the power balance error of the microgrid after correction by the generalized model predictive control algorithm. Figure 5 This is a schematic diagram of the power allocation results of multi-timescale energy management that combines day-ahead planning and real-time scheduling in this embodiment.

[0120] The embodiments of the present invention provide a reference for real-time scheduling by using the optimal power reference value of day-ahead planning, which reduces the computational load of real-time optimization and effectively improves the real-time performance of energy management algorithms. Furthermore, day-ahead planning performs a global search for optimal power allocation on a daily cycle, which can effectively solve for the power allocation with the highest efficiency and lowest operating cost of microgrids under the intermittency of renewable energy.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0122] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. An energy management method applicable to solid oxide fuel cell microgrids, characterized in that, Includes the following steps: Step 1: Establish the optimal control objective function for microgrid energy management. The optimization indicators of this objective function include: economic indicators and durability indicators. The constraint condition for optimizing the objective function is: the photovoltaic power of the photovoltaic cell. Wind turbine power Net power of solid oxide fuel cell (SOFC) system Lithium battery power Hydrogen production equipment power The sum of the load power requirements Phase equilibrium; Step 2: Decompose the solution of the established optimization objective function into two time scales: daily planning with a period of one day and real-time scheduling with a period of one minute. Step 3: For the day-ahead planning timescale of the optimization objective function, use a long short-term memory neural network to calculate the day-ahead power generation of renewable energy. , and load power requirements Perform day-ahead power forecasting; Step 4: For the day-ahead planning timescale of the optimization objective function, based on load power demand... The predicted power of the SOFC system in the next 24 hours is obtained by using a deep deterministic policy gradient reinforcement learning algorithm. By performing a day-ahead forecast, net power is obtained. The current forecast results; Step 5: For the real-time scheduling timescale of the optimization objective function, apply the generalized model predictive control algorithm to the net power. The day-ahead forecast results are scheduled in real time to eliminate errors in the forecast results at each day-ahead scheduling time; In step 1, the objective function for optimal control of microgrid energy management is specifically set as follows: in, This is the first control action in the optimal control strategy set. t is the initial time, and t is the current time. For the scheduling period, The sampling interval is... The number of sampling points included in the scheduling period. As an economic indicator, As a durability indicator, , They are respectively and The preset weighting coefficients, and and The sum of is 1. To optimize the constraints of the objective function; In step 5, the prediction, correction, and optimization steps of the generalized model predictive control algorithm are as follows: (5-1) Power Prediction: Photovoltaic power of photovoltaic cells Wind turbine power Net power of solid oxide fuel cell (SOFC) system Lithium battery power Hydrogen production equipment power The sum of the load power requirements Phase equilibrium; The photovoltaic power at p future moments With fan power and The regression model used for prediction is set as follows: in, For time k to time k+j The predicted value, where p is the forward prediction step size. , For past data step size, express A preset coefficient; The predicted output power of the microgrid is: in, , and They are respectively Always Control power of fuel cells, lithium batteries, and electrolyzers at all times; (5-2) Feedback correction: Correcting prediction errors: In the formula, For the corrected Always Moment The predicted value, where H is a preset constant between 0 and 1, and the predicted disturbance. for The difference between the actual power generation at time k and the predicted power generation at time k: (5-3) Rolling optimization: Using a penalty function, the day-ahead planning reference value is... Integrate this into real-time scheduling to construct a scheduling optimization problem: Where k represents the current time. for The load power requirement at any given time. The corresponding value of the net power of SOFC for the next 24 hours obtained through reinforcement learning; The control quantity is solved based on the scheduling optimization problem to obtain the optimal control power of fuel cell, lithium battery and electrolyzer at several times after the current time k, and the optimal control strategy set is obtained. All the optimal control power at each time after time k is a control action. The control actions in the optimal control strategy set are arranged in ascending order of time, and the first control action in the optimal control strategy set is applied to the microgrid.

2. The method as described in claim 1, characterized in that, In step 1, the power and storage capacity constraints are set as follows: (1) Power range limitation: , , ; (2) Power variation limit: , , ; (3) Storage capacity limit: , ; Where, N i and P r,i These represent the number of devices and the rated output power of the i-th object, respectively. , These are photovoltaic, wind turbine, fuel cell, lithium battery, electrolyzer, and hydrogen storage tank, respectively. SoC represents the battery state of charge, and LoH represents the hydrogen level.

3. The method as described in claim 1, characterized in that, Economic indicators and durability indicators Specifically: ； ； in, Because of the low calorific value of hydrogen, For SOFC efficiency function, This represents the SOFC voltage degradation rate.

4. The method as described in claim 1, characterized in that, Step 2 is as follows: (2-1) Day-ahead planning with a period of 24 hours: The period is 24 hours and the interval is 1 hour. The corresponding parameter in the objective function is set as: T EV =24, Δt=1, N=24; (2-2) Real-time scheduling in minutes: The period is 1 hour, the interval is 5 minutes, and the corresponding parameter in the optimization objective function is set to: T EV =60, Δt=5, N=12.

5. The method as described in claim 1, characterized in that, In step 3, the daily power generation P of renewable energy is... PV P wt and load power demand P load The specific prediction is as follows: Where f() is the output of the LSTM neural network model, For the load power of the next 24 hours, For the prediction input on day d, Let be the feature vector of the t-th hour on day d. eigenvectors The feature representation is set as follows: in, , and The solar radiation, wind speed, and load requirements are given for day d and hour t. , , , , , and These are the forecast data for temperature, dew point temperature, air pressure, humidity, weather type, cloud cover, and visibility for the next t hours.

6. The method as described in claim 3, characterized in that, In step 4, the net power P of the SOFC system for the next 24 hours is calculated using a deep deterministic policy gradient reinforcement learning algorithm. fc The specific details of the day-ahead forecast are as follows: In the deep deterministic policy gradient reinforcement learning algorithm, the actions, states, and rewards are set as follows: Define the action as the net power of SOFC over the next 24 hours: ,in, For the net power in the next t hours, ; Define the state as: ,in, and These represent the initial battery state of charge and initial hydrogen level of the microgrid at time 0 on day d; P d,t For the load power requirement P load The day-ahead predicted power value for the d-th hour; The reward is defined as: .

7. The method as described in claim 1, characterized in that, In step 4, when the net power P is obtained... fc Based on the previous forecast results, the rules for determining the energy of lithium batteries, electrolyzers, and hydrogen storage tanks are set as follows: (1) When the net power P fc When the current forecast result is greater than zero, lithium batteries are charged first, and hydrogen is electrolyzed after they are fully charged. (2) When the power generation of SOFC is equal to the net power P fc When the sum of the previous day's forecast results is less than zero, the power balance is maintained by the discharge of lithium batteries. (3) SOFCs should use electrolyzed hydrogen first, and add hydrogen only when necessary.

Images