Energy storage scheduling method and device of combined cooling heating and power system

Abstract

The invention relates to the technical field of energy system management, and discloses an energy storage scheduling method and device of a combined cooling heating and power system, and the method comprises the following steps: S1, collecting and building a physical model and data of each device of the combined cooling heating and power system; s2, load prediction is carried out by using a long short-term memory (LSTM) neural network; s3, based on a load prediction result, a scheduling optimization objective function is established, and the objective function defines an optimization direction and an evaluation standard; s4, scheduling the charging and discharging processes of the energy storage device by adopting a deep reinforcement learning algorithm; s5, performing real-time scheduling adjustment by using model predictive control; and S6, using rolling horizon optimization to continuously optimize the scheduling strategy on each time step. According to the method, the future load is accurately predicted by adopting the long short-term memory neural network LSTM, and real-time scheduling is realized by combining model prediction control MPC and rolling horizon optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of energy system management, and in particular to an energy storage scheduling method and device for a combined cooling, heating and power system. Background Art

[0002] Combined Cooling, Heating, and Power (CCHP) is an integrated energy system that can provide electricity, heating, and cooling at the same time. It is usually used in scenarios such as buildings and industrial parks that require multiple forms of energy. CCHP systems significantly improve the comprehensive utilization efficiency of energy by effectively utilizing the waste heat of fuels, and are considered to be one of the effective means to achieve low-carbon energy supply and distributed energy utilization. In traditional CCHP systems, the production of electricity, heating, and cooling is often tightly coupled, and cannot flexibly respond to the dynamic changes of various energy loads. The operation of CCHP equipment is usually load-oriented, and the equipment is frequently started and stopped, which not only reduces the service life of the equipment, but also leads to energy waste. With the continuous improvement of modern society's requirements for energy utilization efficiency and environmental protection, how to optimize the operation of CCHP systems has become an important research direction.

[0003] After searching, a method for energy storage scheduling of a combined heat and power system based on a static multi-objective framework was disclosed in the announcement number: CN107370167B. For a combined heat and power system that includes renewable energy, a large-capacity energy storage system can be used to smooth the output fluctuations of renewable energy and cope with uncertain user loads. The present invention is based on a static multi-objective framework and uses the firefly algorithm to calculate the optimal energy storage solution of the combined heat and power system, thereby achieving the optimal economic operation of the combined heat and power system and greatly improving energy utilization.

[0004] However, the above-mentioned methods use static optimization methods to dispatch the energy storage of the combined cooling, heating and power system, but these methods are often unable to cope with rapidly changing load demands and lack sufficient dynamic response capabilities. Static optimization usually shows low scheduling flexibility and high energy waste in actual operation, and cannot give full play to the role of energy storage devices in energy balance. In addition, due to the inability to predict load demand and the lack of accuracy, traditional methods have low energy utilization efficiency in scenarios with large load fluctuations. Summary of the invention

[0005] In order to make up for the above shortcomings, the present invention provides a method and device for energy storage scheduling of a cogeneration system, aiming to improve the traditional static optimization method of the cogeneration system to schedule the energy storage of the cogeneration system, the inability to predict the load demand and the lack of accuracy, often unable to cope with rapidly changing load demands, and lack of sufficient dynamic response capabilities.

[0006] In a first aspect, the present invention provides the following technical solution, a method for energy storage scheduling of a combined cooling, heating and power system, comprising the following steps:

[0007] S1. Collect and establish the physical models and data of each device in the cogeneration system;

[0008] S2, load forecasting by using long short-term memory neural network LSTM to predict the demand for electricity, heating and cooling in future periods;

[0009] S3. Based on the load forecast results, a scheduling optimization objective function is established, where the objective function defines the optimization direction and evaluation criteria, and the objective function includes minimizing operating costs, maximizing energy efficiency, and minimizing carbon emissions;

[0010] S4. Use deep reinforcement learning algorithms to schedule the charging and discharging process of the energy storage device to meet the upper and lower limits of the equipment output, the limitations of the energy storage capacity, and the constraints of power balance;

[0011] S5. During operation, model predictive control is used to make real-time scheduling adjustments based on real-time load and energy storage device status;

[0012] S6. Use rolling horizon optimization to continuously optimize the scheduling strategy at each time step.

[0013] Preferably, in step S1, the operating data of various equipment in the cogeneration system are first collected, including the power output of the power generation equipment, the output power of the heating equipment and the refrigeration equipment, and the charging and discharging status of the battery, the heat storage device and the cold storage device, and then the temperature and humidity of the environment are collected, and finally a physical model of each device is established.

[0014] Preferably, the step S2 includes:

[0015] S201, preprocess the historical load data and environmental data, including normalization, standardizing the data to the [0,1] interval,

[0016] S202, construct and train an LSTM model, with historical load data and environmental data as input and power, heating and cooling demand for multiple time steps in the future as output. The training goal is to minimize the mean square error (MSE), and the Adam optimization algorithm is used to update the model parameters.

[0017] S203: Based on the real-time data input into the trained LSTM model, the power, heat and cooling demand in the future time period is predicted to generate the predicted load value L elec ,L heat ,L cool .

[0018] Preferably, in the S3 step, the operating cost is minimized:

[0019]

[0020] Among them, J 1 is the total operating cost of each device, T is the time range of optimization, including T time steps, C f is the fuel cost of each device, F g is the fuel consumption, C grid is the electricity purchase cost, P grid，t is the power purchased from the grid at time step t, C storage is the operating cost of the energy storage device, P storage,t是 The charge and discharge power of the energy storage device at time step t, where a positive value indicates discharge and a negative value indicates charge.

[0021] Maximize energy efficiency:

[0022]

[0023] Among them, J 2 is the energy efficiency, P out，t is the power output of the system at time step t, Q out，t is the heat output of the system at time step t, T is the time range of optimization, including T time steps, F g is the fuel consumption, which indicates the amount of fuel used by the system in time step t, P grid,t is the power purchased from the grid at time step t;

[0024] Minimize Carbon Emissions:

[0025]

[0026] Among them, J 3 is the carbon emission target, the goal is to minimize it, e g is the carbon emission factor of the fuel, which indicates the carbon emission generated per unit of fuel consumption, T is the optimization time range, including T time steps, F g is the fuel consumption, which represents the amount of fuel used by the system in time step t.

[0027] The final comprehensive objective function is:

[0028] J min =α 1 J 1 -α 2 J 2 +α 3 J 3

[0029] J min: Minimize the comprehensive objective function, which represents the weighted sum of different objectives, J 1 is the operating cost target, J 2 is the energy efficiency target, J 3 is the carbon emission target. 1 is the weight coefficient used to balance the importance of operating costs, α 2 is the weight coefficient used to balance the importance of energy efficiency, α 3 is the weight coefficient used to balance the importance of carbon emissions;

[0030] Constraint setting: including upper and lower limits of equipment output, energy storage capacity restrictions, and power balance constraints:

[0031] P min，i ≤P i，t ≤P max，i

[0032] 0≤E storage,t ≤C storage

[0033] P gen,t +P storage,t =L elec,t +P grid,t

[0034] P min,i ,P max,i are the minimum and maximum power output limits of device i;

[0035] P i,t is the power output of device i at time step t

[0036] E storage,t is the energy state of the energy storage device at time step t;

[0037] C storage is the maximum capacity of the energy storage device;

[0038] P gen,t is the electrical power output by the power generation equipment at time step t;

[0039] P storage,t is the power output of the energy storage device at time step t. If the value is positive, it means that the energy storage device is discharging to supply the load; if the value is negative, it means that the energy storage device is charging;

[0040] L elec,t is the power load demand at time step t, which indicates the amount of electricity required by the user at this point in time;

[0041] P grid,t is the power purchased from the external grid at time step t.

[0042] Preferably, in the step S4, the specific process includes:

[0043] S401, State space and action space definition

[0044] State space s t :The state space defines the state of the cogeneration system at a certain moment, which is used to describe the complete operation of the cogeneration system. The state contains the following variables:

[0045] Energy storage state E storage,t : Indicates the current energy level of the energy storage device;

[0046] Load Forecast elec,t ,L heat,t ,L cool,t : Future load demand prediction by LSTM;

[0047] Current device status P gen,t , P grid,t : Including the output power of power generation equipment and external power purchase power;

[0048] Environmental parameters T, H: current ambient temperature and humidity;

[0049] Action Space a t : The action space defines the charging and discharging operations of the energy storage device:

[0050] Energy storage charging and discharging power P storage,t : Charging is a positive value and discharging is a negative value, indicating the charging and discharging behavior of the energy storage device at time step t;

[0051] S402. Reward function design

[0052] R t =-(α 1 J 1 +α 2 J 2 +α 3 J 3 )

[0053] Among them, R t is the reward value at time step t, which is used to guide the learning direction of the deep reinforcement learning algorithm, α 1 is the weight coefficient used to balance the importance of operating costs, J 1 is the operating cost target, α 2 is the weight coefficient used to balance the importance of energy efficiency, J 2 is the energy efficiency target, α 3 is the weight coefficient used to balance the importance of carbon emissions, J 3 is the carbon emission target;

[0054] Minimize operating costs 1 : Includes fuel, electricity purchase and energy storage operating costs, increasing incentives by reducing operating costs;

[0055] Maximize energy efficiency 2 : Increase reward value by using energy efficiently;

[0056] Minimizing carbon emissions 3 : Increase rewards by reducing carbon emissions;

[0057] S403. Deep reinforcement learning training process

[0058] Initialize the policy network and value network:

[0059] Policy network: used to output in state s t The next action a t , i.e. how the energy storage device should be charged or discharged;

[0060] Value network: used to evaluate the t Next, perform action a t The expected return Q(s) t ,a t ), the goal of the Critic network is to help the Actor network improve its strategy;

[0061] Experience replay: When the system interacts with the environment, the data of each interaction is stored in the experience replay buffer:

[0062] (s t , α t , R t ,s t+1 )

[0063] The network is trained by sampling small batches of data from the experience replay buffer to avoid correlation between data and improve the stability and efficiency of training;

[0064] Target network soft update:

[0065] In order to improve the stability of network training, the target network is used for updating. The parameters of the target network are obtained by soft updating of the policy network and the value network. The update rule is:

[0066] θ′=τθ+(1-τ)θ′

[0067] Among them, θ′ represents the parameters of the target network, θ represents the parameters of the main network, and τ is 0.001, which is used to control the update speed.

[0068] Policy improvement: Use the gradient ascent method to update the policy network, the goal is to increase the policy in state s t Next, perform action a t The return:

[0069]

[0070] is the gradient of the policy, that is, the update direction of the policy function;

[0071] is the state s t Expected value;

[0072] It is the gradient of the action value Q function with respect to action a, which indicates the rate of change of the action value function under state s and action a;

[0073] s=s t , a=μ(s t ) refers to the action selected by the current policy network μ in state st;

[0074] is the gradient of the policy function μ with respect to its parameter θμ.

[0075] The value network update uses gradient descent to minimize the value network's loss function, which is the mean squared error between the actual return and the target Q value:

[0076]

[0077] Among them, L(θ Q ) is the loss function, which is used to measure the estimation error of the Q function;

[0078] θ Q are the parameters of the action-value Q function;

[0079] E(s t ,a t ,R t ,s t+1 ) is the expected value of a set of sample data, and the expectation is based on the sampling in the experience pool (s t ,a t ,R t ,s t+1 ) quaternion;

[0080] Q(s t ,a t ∣θ Q ) is the estimated Q value of the state st and action at given by the current Q function;

[0081] y tis the target Q value, used to train the Q function:

[0082] y t =R t +γQ′(s t+1 ,μ′(s t+1 )|θ Q′ )

[0083] R t It’s an instant reward;

[0084] γ is a discount factor, which indicates the importance of future rewards;

[0085] Q′(s t+1 ,μ′(s t+1 )|θ Q′ ) is the next state s given by the target Q network t+1 The estimated value of is calculated using the target Q network and the target policy network;

[0086] θ Q′ are the parameters of the target Q network.

[0087] S404. Execution of energy storage optimization scheduling

[0088] Action selection:

[0089] At each time step t, use the policy network based on the current state s t Select action a t , which determines the charging and discharging power of the energy storage device at the current moment;

[0090] Action a t Will directly affect the energy state E of the energy storage device storage,t+1 , thus affecting future energy supply and demand;

[0091] Perform charge and discharge operations:

[0092] According to the selected action at, the corresponding charging and discharging operations are performed, the energy state of the energy storage device is changed, and the future operation state is affected;

[0093] Interact with the environment and give feedback:

[0094] After executing an action, the system interacts with the environment and generates a new state s t+1 and reward R t， These new states will serve as input for the next decision, enter the experience replay buffer, and be used for further strategy improvement.

[0095] Preferably, in step S5, the optimal strategy learned by deep reinforcement learning during the training process will be used as the initial plan for real-time scheduling control. During actual operation, model predictive control will optimize and adjust the strategy provided by deep reinforcement learning in real time based on the current state and load forecast to cope with dynamic changes in the environment.

[0096] Preferably, the step S6 includes:

[0097] S601, load forecast update: at each time step, use real-time data to input the LSTM model to obtain new future load forecast values, which will affect the next dispatch decision;

[0098] S602, optimization solution update: Combine the updated load forecast and the current state, use the strategy trained by deep reinforcement learning to optimize the energy storage equipment and power generation equipment, and recalculate the optimal charging and discharging plan and power output at each time step;

[0099] S603, control strategy execution: the optimal control strategy obtained by solving is executed at each time step, and refined and adjusted through model predictive control.

[0100] In a second aspect, the present invention provides the following technical solution: an energy storage scheduling device for a combined cooling, heating and power system, comprising:

[0101] The CCHP equipment modeling module is used to model the power generation equipment, heating equipment and refrigeration equipment in the CCHP system. The model covers the output power, energy efficiency characteristics and operating condition parameters of the equipment; it is used to establish the model of the energy storage device, including batteries, heat storage and cold storage devices, and describe the charging and discharging efficiency, capacity and response speed characteristics of the energy storage device;

[0102] The load forecasting module is used to forecast the cooling, heating and electricity loads in the future period, providing a basis for the formulation of dispatching strategies. The forecast is based on historical load data and environmental factors.

[0103] The objective function module is used to determine the scheduling objective function according to the operation requirements and strategies. The objective function contains multiple optimization objectives, including economy, energy efficiency and environmental impact;

[0104] The energy storage optimization scheduling module is used to optimize the charging and discharging of the energy storage device based on the objective function. The optimization algorithm is used to achieve the objective function requirements while satisfying the constraints.

[0105] Real-time dispatch control module, used to dispatch and adjust charging and discharging and cooling and heating production in real time based on real-time load and energy storage status to adapt to dynamically changing load demands;

[0106] The rolling horizon optimization module is used to continuously optimize the control strategy at each time step.

[0107] In a third aspect, the invention provides the following technical solution: a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the energy storage scheduling method for the above-mentioned cogeneration system when executing the computer program.

[0108] In a fourth aspect, the present invention provides the following technical solution: a readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the energy storage scheduling method of the above-mentioned cogeneration system.

[0109] The beneficial effects of the present invention are as follows:

[0110] 1. In the present invention, a long short-term memory neural network (LSTM) is used to accurately predict future loads, and model predictive control (MPC) and rolling time domain optimization are combined to achieve real-time scheduling. This method can effectively cope with load fluctuations, perform charging and discharging operations of the energy storage device in advance, and make reasonable energy distribution between load peaks and troughs, significantly improving energy utilization efficiency. Compared with traditional methods, the method of the present invention greatly reduces energy waste and achieves a significant improvement in energy utilization efficiency through accurate prediction and real-time adjustment of loads. Experimental results show that energy utilization can be increased to 82%.

[0111] 2. The method of the present invention can accurately predict the future demand for electricity, heat and cooling through the long short-term memory neural network (LSTM), and optimize the scheduling in real time through the model predictive control (MPC). Energy storage charging is carried out during the low electricity price period, and the energy of the energy storage device is released during the peak load period to meet the demand, which greatly reduces the operating cost of the system. The comparative experimental results show that the operating cost is reduced by 15%.

[0112] 3. The combination of load prediction and rolling time domain optimization in the present invention significantly improves the system's ability to cope with load fluctuations. By predicting load changes through LSTM and using MPC for rolling optimization scheduling, the system can quickly adjust when the load fluctuates, and the response speed is improved to within 5 minutes. Compared with the traditional static scheduling method, the present invention has a faster response speed and is more flexible, and can effectively cope with rapid changes in power, heat and cooling loads. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 A method flow chart of the energy storage scheduling method for the combined cooling, heating and power system proposed by the present invention;

[0114] Figure 2This is a logical architecture block diagram of the energy storage scheduling device for the combined cooling, heating and power system proposed in the present invention. DETAILED DESCRIPTION

[0115] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in the present invention are only part of the embodiments of the present invention, not all of the embodiments. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0116] Embodiment 1:

[0117] Reference Figure 1 In a first embodiment of the present invention, the present invention provides an energy storage scheduling method for a combined cooling, heating and power system, comprising the following steps:

[0118] S1. Collect and establish the physical models and data of each device in the cogeneration system;

[0119] S2, load forecasting by using long short-term memory neural network LSTM to predict the demand for electricity, heating and cooling in future periods;

[0120] S3. Based on the load forecast results, establish the dispatch optimization objective function, which defines the optimization direction and evaluation criteria. The objective function includes minimizing operating costs, maximizing energy efficiency and minimizing carbon emissions.

[0121] S4. Use deep reinforcement learning algorithms to schedule the charging and discharging process of the energy storage device to meet the upper and lower limits of the equipment output, the limitations of the energy storage capacity, and the constraints of power balance;

[0122] S5. During operation, model predictive control is used to make real-time scheduling adjustments based on real-time load and energy storage device status;

[0123] S6. Use rolling horizon optimization to continuously optimize the scheduling strategy at each time step.

[0124] Specifically, S1. Collect and establish the physical model and data of each device in the cogeneration system

[0125] effect:

[0126] Collect and establish physical models and data of various equipment in the system, including the status of power generation equipment, heating equipment, refrigeration equipment, batteries, thermal storage devices and cold storage devices, as well as environmental parameters (such as temperature and humidity):

[0127] Providing the system's physical model and operating data lays the foundation for subsequent load forecasting, optimal scheduling and real-time control, ensuring the model accuracy and scheduling feasibility of the entire system.

[0128] S2. Load forecasting using long short-term memory neural network (LSTM)

[0129] effect:

[0130] By inputting historical load data and environmental data, the LSTM model is used to predict the load of electricity, heating and cooling demand in future periods.

[0131] Efficiency optimization provides predicted load demand data, enabling the system to foresee future cooling, heating and electricity demand in advance and make reasonable scheduling decisions.

[0132] Improve the system's adaptability to future uncertain demands and reduce energy waste caused by mismatch between supply and demand.

[0133] S3. Establishing the dispatch optimization objective function based on the load forecast results

[0134] effect:

[0135] According to the results of load forecasting, the system's scheduling optimization objective function is established. The objective function includes minimizing operating costs, maximizing energy utilization efficiency, minimizing carbon emissions, etc.

[0136] Effect:

[0137] It clarifies the adjustments and guides the subsequent energy storage optimization and real-time scheduling to ensure that the overall goals of the system can be achieved.

[0138] The diversity of optimization objectives ensures that the system not only pursues economy during operation, but also takes into account energy efficiency and environmental protection.

[0139] S4. Using Deep Reinforcement Learning (DRL) Algorithm for Scheduling Energy Storage Devices

[0140] effect:

[0141] Using deep reinforcement learning algorithms, the charging and discharging of energy storage devices are optimized and scheduled to ensure that the upper and lower limits of equipment output, energy storage capacity limitations, and power balance constraints are met.

[0142] Effect:

[0143] By learning the optimal charging and discharging fluctuations of the energy storage device, we can make full use of the energy storage to balance the demand for cooling, heating and electricity, reduce peak loads and fill valleys, and reduce the cost of purchasing electricity.

[0144] Through the self-learning ability of deep reinforcement learning, the system's scheduling strategy can be continuously optimized to adapt to the dynamically changing environment.

[0145] S5. Real-time scheduling adjustment based on model predictive control (MPC)

[0146] effect:

[0147] During actual operation, model predictive control (MPC) is used to make real-time scheduling adjustments to energy storage devices and power generation equipment based on real-time load and energy storage device status.

[0148] Effect:

[0149] It achieves a rapid response to environmental changes, so that the scheduling results can still maintain efficient operation even when the load is negative.

[0150] MPC is used to adjust the strategies learned by deep reinforcement learning in real time to optimize the system operation in the current state and ensure the real-time and flexibility of the scheduling strategy.

[0151] S6. Use rolling horizon optimization to continuously optimize the scheduling strategy at each time step

[0152] effect:

[0153] The rolling time domain optimization method is used to continuously optimize the scheduling strategy of energy storage and power generation equipment according to the latest load forecast and system status at each time step.

[0154] Effect:

[0155] It ensures that the system can make the best decision based on the latest information at each time step, making the scheduling smooth.

[0156] The dynamic adaptability and overall operating efficiency of the system have been improved, ensuring that the combined cooling, heating and power system is always in the optimal operating state in a complex environment.

[0157] In step S1, the operating data of various equipment in the cogeneration system are first collected, including the power output of the power generation equipment, the output power of the heating equipment and the refrigeration equipment, as well as the charging and discharging status of the battery, heat storage device and cold storage device. Then the temperature and humidity of the environment are collected, and finally the physical model of each device is established.

[0158] The S2 step includes:

[0159] S201, preprocess the historical load data and environmental data, including normalization, standardizing the data to the [0,1] interval,

[0160] S202, construct and train an LSTM model, with historical load data and environmental data as input and power, heating and cooling demand for multiple time steps in the future as output. The training goal is to minimize the mean square error (MSE), and the Adam optimization algorithm is used to update the model parameters.

[0161] S203: Based on the real-time data input into the trained LSTM model, the power, heat and cooling demand in the future time period is predicted to generate the predicted load value L elec ,L heat ,L cool .

[0162] Specifically,

[0163] Data preprocessing:

[0164] The collected historical load data and environmental data are preprocessed, including normalization, so that the model can effectively learn the relationship between the data.

[0165] Split the data into training and testing sets to train a Long Short-Term Memory (LSTM) neural network model.

[0166] LSTM model training and load forecasting:

[0167] Use LSTM models to train historical data to predict future electricity, heating and cooling demand.

[0168] The LSTM model obtains the load change trend by learning the long-term and short-term dependencies in the time series.

[0169] Prepare in advance: Forecast future load demand so that the system can prepare for upcoming demand changes in advance and reduce the adverse effects of imbalance in energy supply and demand.

[0170] Better meet demand: By accurately predicting the demand for cooling, heating and electricity, the system can precisely dispatch energy storage devices and power generation equipment to ensure that energy supply meets actual demand and avoid oversupply or undersupply.

[0171] Load management: Load forecasting results will directly affect the charging and discharging decisions of energy storage devices. By predicting load changes in advance, charging can be done during low periods and discharging during peak periods, thereby achieving peak shaving and valley filling.

[0172] In step S3, minimize the running cost:

[0173]

[0174] Among them, J 1 is the total operating cost of each device, T is the time range of optimization, including T time steps, C f is the fuel cost of each device, F g is the fuel consumption, C grid is the electricity purchase cost, P grid，t is the power purchased from the grid at time step t, C storage is the operating cost of the energy storage device, Pstorage,t是 The charge and discharge power of the energy storage device at time step t, where a positive value indicates discharge and a negative value indicates charge.

[0175] Maximize energy efficiency:

[0176]

[0177] Among them, J 2 is the energy efficiency, P out，t is the power output of the system at time step t, Q out，t is the heat output of the system at time step t, T is the time range of optimization, including T time steps, F g is the fuel consumption, which indicates the amount of fuel used by the system in time step t, P grid,t is the power purchased from the grid at time step t;

[0178] Minimize Carbon Emissions:

[0179]

[0180] Among them, J 3 is the carbon emission target, the goal is to minimize it, e g is the carbon emission factor of the fuel, which indicates the carbon emission generated per unit of fuel consumption, T is the optimization time range, including T time steps, F g is the fuel consumption, which represents the amount of fuel used by the system in time step t.

[0181] The final comprehensive objective function is:

[0182] J min =α 1 J 1 -α 2 J 2 +α 3 J 3

[0183] J min : Minimize the comprehensive objective function, which represents the weighted sum of different objectives, J 1 is the operating cost target, J 2 is the energy efficiency target, J 3 is the carbon emission target. 1 is the weight coefficient used to balance the importance of operating costs, α 2 is the weight coefficient used to balance the importance of energy efficiency, α 3 is the weight coefficient used to balance the importance of carbon emissions;

[0184] Constraint setting: including upper and lower limits of equipment output, energy storage capacity restrictions, and power balance constraints:

[0185] P min，i ≤P i，t ≤P max，i

[0186] 0≤E storage,t ≤C storage

[0187] P gen,t +P storage,t =L elec,t +P grid,t

[0188] P min,i ,P max,i are the minimum and maximum power output limits of device i;

[0189] P i,t is the power output of device i at time step t

[0190] E storage,t is the energy state of the energy storage device at time step t;

[0191] C storage is the maximum capacity of the energy storage device;

[0192] P gen,t is the electrical power output by the power generation equipment at time step t;

[0193] P storage,t is the power output of the energy storage device at time step t. If the value is positive, it means that the energy storage device is discharging to supply the load; if the value is negative, it means that the energy storage device is charging;

[0194] L elec,t is the power load demand at time step t, which indicates the amount of electricity required by the user at this point in time;

[0195] P grid,t is the power purchased from the external grid at time step t.

[0196] Specifically, the objective function is defined as:

[0197] Establish objective functions, including minimizing operating costs, maximizing energy efficiency, and minimizing carbon emissions.

[0198] Operating costs include the cost of fuel consumption, the cost of purchasing electricity from the grid, and the cost of charging and discharging energy storage devices.

[0199] Energy efficiency is maximized by reducing fuel consumption, while energy waste is reduced by optimizing the coordinated operation of equipment.

[0200] Target weight setting:

[0201] Set weight coefficients for each goal, dynamically adjust the priority of different goals according to system requirements, and achieve flexible multi-objective optimization.

[0202] Effect:

[0203] The system scheduling direction is clear: through the establishment of the objective function, clear optimization goals and evaluation criteria are provided for scheduling optimization, ensuring that the scheduling strategy is in line with the overall interests of the system.

[0204] Multi-objective balance: The weight setting of the objective function ensures that the optimization direction can be flexibly adjusted under different operating scenarios, such as giving priority to reducing costs during peak load periods and giving priority to reducing carbon emissions in low-carbon operating scenarios.

[0205] Economical and environmentally friendly: By optimizing operating costs and carbon emissions, the system can not only reduce economic costs during operation, but also reduce the negative impact on the environment.

[0206] In step S4, the specific process includes:

[0207] S401, State space and action space definition

[0208] State space s t :The state space defines the state of the cogeneration system at a certain moment, which is used to describe the complete operation of the cogeneration system. The state contains the following variables:

[0209] Energy storage state E storage,t : Indicates the current energy level of the energy storage device;

[0210] Load Forecast elec,t ,L heat,t ,L cool,t : Future load demand prediction by LSTM;

[0211] Current device status P gen,t , P grid,t : Including the output power of power generation equipment and external power purchase power;

[0212] Environmental parameters T, H: current ambient temperature and humidity;

[0213] Action Space a t : The action space defines the charging and discharging operations of the energy storage device:

[0214] Energy storage charging and discharging power P storage,t : Charging is a positive value and discharging is a negative value, indicating the charging and discharging behavior of the energy storage device at time step t;

[0215] S402. Reward function design

[0216] Rt =-(α 1 J 1 +α 2 J 2 +α 3 J 3 )

[0217] Among them, R t is the reward value at time step t, which is used to guide the learning direction of the deep reinforcement learning algorithm, α 1 is the weight coefficient used to balance the importance of operating costs, J 1 is the operating cost target, α 2 is the weight coefficient used to balance the importance of energy efficiency, J 2 is the energy efficiency target, α 3 is the weight coefficient used to balance the importance of carbon emissions, J 3 is the carbon emission target;

[0218] Minimize operating costs 1 : Includes fuel, electricity purchase and energy storage operating costs, increasing incentives by reducing operating costs;

[0219] Maximize energy efficiency 2 : Increase reward value by using energy efficiently;

[0220] Minimizing carbon emissions 3 : Increase rewards by reducing carbon emissions;

[0221] S403. Deep reinforcement learning training process

[0222] Initialize the policy network and value network:

[0223] Policy network: used to output in state s t The next action a t , i.e. how the energy storage device should be charged or discharged;

[0224] Value network: used to evaluate the state s t Next, perform action a t The expected return Q(s) t ,a t ), the goal of the Critic network is to help the Actor network improve its strategy;

[0225] Experience replay: When the system interacts with the environment, the data of each interaction is stored in the experience replay buffer:

[0226] (s t , a t , R t ,s t+1 )

[0227] The network is trained by sampling small batches of data from the experience replay buffer to avoid correlation between data and improve the stability and efficiency of training;

[0228] Target network soft update:

[0229] In order to improve the stability of network training, the target network is used for updating. The parameters of the target network are obtained by soft updating of the policy network and the value network. The update rule is:

[0230] θ′=τθ+(1-τ)θ′

[0231] Among them, θ′ represents the parameters of the target network, θ represents the parameters of the main network, and τ is 0.001, which is used to control the update speed.

[0232] Policy improvement: Use the gradient ascent method to update the policy network, the goal is to increase the policy in state s t Next, perform action a t The return:

[0233]

[0234] is the gradient of the policy, that is, the update direction of the policy function;

[0235] is the state s t Expected value;

[0236] It is the gradient of the action value Q function with respect to action a, which indicates the rate of change of the action value function under state s and action a;

[0237] s=s t , a=μ(s t ) refers to the action selected by the current policy network μ in state st;

[0238] is the gradient of the policy function μ with respect to its parameter θμ.

[0239] The value network update uses gradient descent to minimize the value network's loss function, which is the mean squared error between the actual return and the target Q value:

[0240]

[0241] Among them, L(θ Q ) is the loss function, which is used to measure the estimation error of the Q function;

[0242] θ Q are the parameters of the action-value Q function;

[0243] E(s t ,a t ,R t ,s t+1 ) is the expected value of a set of sample data, and the expectation is based on the sampling in the experience pool (s t ,a t ,R t ,s t+1 ) quaternion;

[0244] Q(s t ,a t ∣θ Q ) is the estimated Q value of the state st and action at given by the current Q function;

[0245] y t is the target Q value, used to train the Q function:

[0246] y t =R t +γQ′(s t+1 ,μ′(s t+1 )|θ Q′ )

[0247] R t It’s an instant reward;

[0248] γ is a discount factor, which indicates the importance of future rewards;

[0249] Q′(s t+1 ,μ′(s t+1 )|θ Q′ ) is the next state s given by the target Q network t+1 The estimated value of is calculated using the target Q network and the target policy network;

[0250] θ Q′ are the parameters of the target Q network.

[0251] S404. Execution of energy storage optimization scheduling

[0252] Action selection:

[0253] At each time step t, use the policy network based on the current state s t Select action a t , which determines the charging and discharging power of the energy storage device at the current moment;

[0254] Action a t Will directly affect the energy state E of the energy storage device storage,t+1 , thus affecting future energy supply and demand;

[0255] Perform charge and discharge operations:

[0256] According to the selected action at, the corresponding charging and discharging operations are performed, the energy state of the energy storage device is changed, and the future operation state is affected;

[0257] Interact with the environment and give feedback:

[0258] After executing an action, the system interacts with the environment and generates a new state s t+1 and reward R t , these new states will serve as input for the next decision, enter the experience replay buffer, and be used for further strategy improvement.

[0259] Specifically, the state space and action space are defined as:

[0260] The state space includes the charging and discharging state of the energy storage device, the predicted load demand, the operating state of the power generation equipment, and environmental conditions.

[0261] The action space defines the charging and discharging behaviors of the energy storage device, and the goal is to select appropriate operations based on the load forecast and current status.

[0262] Deep reinforcement learning training:

[0263] Using deep reinforcement learning (such as DDPG), the optimal scheduling strategy under different states is learned through the interaction of the policy network (Actor) and the value network (Critic).

[0264] The reward function is set based on the objective function, and through the iteration of reinforcement learning, the optimal charging and discharging strategy is found to minimize operating costs, improve energy efficiency and reduce carbon emissions.

[0265] Peak shaving and valley filling: By learning from historical experience, energy storage devices can be charged when electricity demand is low and discharged when it is peak, thereby achieving load shaving and valley filling and reducing the pressure on power generation equipment.

[0266] Strong adaptability: Deep reinforcement learning makes the scheduling strategy highly adaptable. It can automatically learn and adjust the strategy according to the changing load and environmental conditions, thereby improving the operational robustness of the system.

[0267] Improve energy efficiency: The optimal strategy learned through reinforcement learning can enable energy storage devices and power generation equipment to operate in coordination, maximizing overall energy efficiency and reducing electricity purchase costs.

[0268] In step S5, the optimal strategy learned by deep reinforcement learning during the training process will be used as the initial plan for real-time scheduling control. During actual operation, model predictive control will optimize and adjust the strategy provided by deep reinforcement learning in real time based on the current state and load forecast to cope with dynamic changes in the environment.

[0269] Specifically,

[0270] Real-time status collection and update:

[0271] During system operation, the current status data of the equipment (such as energy storage status, power demand, environmental data, etc.) is collected in real time.

[0272] Based on the real-time status, model predictive control (MPC) is used to predict the future system state at each time step and determine the optimal control strategy in the short term.

[0273] Adjust strategies in real time:

[0274] Using the strategies learned from deep reinforcement learning as the basis and combining it with MPC to achieve real-time adjustment, the scheduling strategy can be continuously refined and optimized according to actual conditions during operation.

[0275] Adapting to rapid changes: MPC provides a mechanism for fine-tuning in a rapidly changing environment, ensuring that the system maintains optimal scheduling when load fluctuates and the external environment changes.

[0276] Improve real-time performance: Through real-time data collection and model prediction, it is possible to respond quickly to changes in system status, reduce the negative impact of load fluctuations, and maximize system stability and operating efficiency.

[0277] Maintain optimal operation: Real-time adjustments enable energy storage devices and power generation equipment to be adjusted at any time based on current demand and future forecasts, avoiding energy waste and ensuring optimal operation.

[0278] Step S6 includes:

[0279] S601, load forecast update: at each time step, use real-time data to input the LSTM model to obtain new future load forecast values, which will affect the next dispatch decision;

[0280] S602, optimization solution update: Combine the updated load forecast and the current state, use the strategy trained by deep reinforcement learning to optimize the energy storage equipment and power generation equipment, and recalculate the optimal charging and discharging plan and power output at each time step;

[0281] S603, control strategy execution: the optimal control strategy obtained by solving is executed at each time step, and refined and adjusted through model predictive control.

[0282] Specifically, rolling forecast and optimization:

[0283] At each time step, the optimal scheduling plan is continuously recalculated using rolling horizon optimization, combining the new load data predicted by LSTM and the current system status.

[0284] Each optimization period contains forecast information for multiple future time steps, ensuring that scheduling decisions take into account medium- and long-term demand changes.

[0285] Dynamic update of scheduling strategy:

[0286] The optimization results are updated at each time step, and rolling optimization is performed using real-time data and load forecast results to ensure that the dispatch plan is always adjusted based on the latest data.

[0287] Continuous optimization: Rolling time domain optimization ensures that the system is continuously optimized during operation, and continuously adjusts the scheduling strategies of energy storage and power generation equipment to maximize the overall energy efficiency of the system.

[0288] Flexible response to changes: When load demand, energy prices or external conditions change, rolling time domain optimization can respond promptly to new changes to ensure the effectiveness of the scheduling strategy in the new environment.

[0289] Efficient operation: Through rolling optimization, the decision at each time step is the optimal solution based on the latest information and predictions, avoiding the failure of scheduling strategies and ensuring efficient and sustainable operation.

[0290] In order to verify the effectiveness of the method of the present invention, the following example selects a combined heat and power system in a certain area for scheduling optimization, focusing on how the load forecasting and rolling time domain optimization in the present invention are specifically applied to the energy storage scheduling of the system, so as to achieve efficient energy utilization and reduce operating costs.

[0291] 1.1. System parameters and load conditions

[0292] System parameters:

[0293] Power generation equipment: Gas turbine, rated power 2000kW, efficiency 35%.

[0294] Thermal storage equipment: The thermal storage capacity is 5000kWh and the thermal efficiency is 90%.

[0295] Cold storage equipment: The cold storage capacity is 4000kWh and the cooling efficiency is 85%.

[0296] Battery energy storage equipment: capacity is 1000kWh, charging and discharging efficiency is 95%.

[0297] Load data:

[0298] The load data of a certain day in summer is selected as an example. The load on that day shows obvious peak-to-valley changes, with the peak value of the power load being 1800kW and the valley value being 800kW. The peak value of the heat load is 1500kW and the peak value of the cold load is 1000kW.

[0299] 2. Load forecasting process

[0300] The long short-term memory (LSTM) neural network is used to predict future load demand. The specific process is as follows:

[0301] Input data: Use the load data of the past 7 days, the ambient temperature and humidity data of the current day.

[0302] Forecasting goal: To predict the demand for electricity, heating and cooling every 15 minutes in the next 24 hours.

[0303] After training, the LSTM network outputs the load demand curve for the next 24 hours. The forecast results show that the peak load of the day occurs between 12:00 and 14:00, and the trough load occurs between 2:00 and 5:00 in the morning.

[0304] 3. Scheduling Optimization Process

[0305] Based on the predicted load data, model predictive control (MPC) is used to perform rolling horizon optimization scheduling. The optimization objectives are:

[0306] Minimize operating costs.

[0307] Maximize energy efficiency.

[0308] Minimize carbon emissions.

[0309] Optimization steps:

[0310] Initialize the current state:

[0311] Get the system's current battery energy state (800kWh), thermal storage energy state (3000kWh), and cold storage energy state (2000kWh).

[0312] Read the current load demand: power load 1500kW, heat load 1200kW, cooling load 800kW.

[0313] Predicting future loads:

[0314] The LSTM model is used to predict the load demand in the next 4 hours and obtain the prediction curves of electricity, heating and cooling loads.

[0315] Set up the optimization problem:

[0316] Based on the objective function and constraints, a scheduling optimization model is established. The objective function is:

[0317]

[0318] in:

[0319] C op,t : Operating costs, including fuel and grid electricity purchase costs.

[0320] η sys,t : Energy efficiency of the system.

[0321] e co2,t : Carbon emissions.

[0322] α 1 ,α 2 ,α 3 : Weight factor.

[0323] Solve for the optimal control input:

[0324] The rolling time domain optimization method is used to achieve the optimal match of load demand by optimizing the charging and discharging power of the battery, the charging power of the heat storage device, the charging power of the cold storage device and the output power of the power generation equipment.

[0325] Execution schedule:

[0326] During the peak period of 12:00-14:00, the power generation equipment is used for full load output, while the thermal storage and battery energy are released to meet the load demand.

[0327] During the off-peak period from 2:00 a.m. to 5:00 a.m., the battery is charged using the low-peak electricity price of the power grid, and heat and cold energy are stored in the heat storage and cold storage devices to prepare for the subsequent load peak.

[0328] 4. Comparison of experimental results with existing technologies

[0329] In order to verify the effectiveness of the method of the present invention, several existing technologies were selected for comparison:

[0330] Existing technology:

[0331] CN107370167B, an energy storage scheduling method for a combined cooling, heating and power system based on a static multi-objective framework.

[0332] "A Multi-Objective Optimization Strategy for CHP Systems with Battery Storage", published in the journal "Energy", proposed a multi-objective optimization method, but it adopted a static optimization method without rolling prediction and real-time adjustment.

[0333] Table 1 Comparison of experimental results of existing technologies

[0334]

[0335]

[0336] Comparative experimental results:

[0337] Energy efficiency:

[0338] The method of the present invention realizes early response and real-time adjustment to the demand for electricity, heat and cooling power through load forecasting and rolling time domain optimization, so that the energy utilization efficiency of the system reaches 82%.

[0339] Existing technology 1: Based on static scheduling, the energy utilization efficiency is 70%. Due to the failure to respond to load fluctuations in real time, there is energy waste.

[0340] Existing technology 2: adopts static optimization strategy, and the energy utilization efficiency is 73%.

[0341] Running costs:

[0342] The method of the present invention: through rolling time domain optimization and valley charging, the operation cost of the system is reduced by 15%, from 5000 yuan / day to 4250 yuan / day.

[0343] Existing technology 1: Due to the failure to reasonably utilize off-peak electricity prices, the operating cost is 4,800 yuan / day.

[0344] Existing technology 2: Although multi-objective optimization was performed, the operating cost was 4,600 yuan / day due to the lack of real-time adjustment.

[0345] Response speed:

[0346] The method of the present invention: load prediction and real-time optimization significantly improve the load response speed of the system and can adjust the load fluctuation within 5 minutes.

[0347] Existing technology 1: adopts static scheduling mode, which has slow response speed and adjustment time of 15 minutes.

[0348] Prior Art 2: Due to the lack of real-time prediction, the adjustment time exceeds 20 minutes.

[0349] 5. Analysis of the advantages of the method of the present invention

[0350] Accurate load forecasting: Through the accurate prediction of future loads by the long short-term memory neural network (LSTM), the system can schedule charging and discharging in advance, reduce emergency responses during peak hours, and improve system stability.

[0351] Real-time adjustment of rolling horizon optimization: Different from static optimization, the method of the present invention can adjust the energy storage scheduling strategy in real time according to the actual operating status of the system and environmental changes through model predictive control (MPC) and rolling horizon optimization. This greatly improves the operating efficiency of the system and reduces energy waste caused by unpredictable load fluctuations.

[0352] Economical efficiency of peak shaving and valley filling: The method of the present invention utilizes low-priced electricity from the power grid to store energy during low-load periods, and releases the energy of the energy storage device during peak load periods, which not only balances the load but also significantly reduces operating costs.

[0353] The beneficial effects of the present invention are as follows:

[0354] 1. In the present invention, a long short-term memory neural network (LSTM) is used to accurately predict future loads, and model predictive control (MPC) and rolling time domain optimization are combined to achieve real-time scheduling. This method can effectively cope with load fluctuations, perform charging and discharging operations of the energy storage device in advance, and make reasonable energy distribution between load peaks and troughs, significantly improving energy utilization efficiency. Compared with traditional methods, the method of the present invention greatly reduces energy waste and achieves a significant improvement in energy utilization efficiency through accurate prediction and real-time adjustment of loads. Experimental results show that energy utilization can be increased to 82%.

[0355] 2. The method of the present invention can accurately predict the future demand for electricity, heat and cooling through the long short-term memory neural network (LSTM), and optimize the scheduling in real time through the model predictive control (MPC). Energy storage charging is carried out during the low electricity price period, and the energy of the energy storage device is released during the peak load period to meet the demand, which greatly reduces the operating cost of the system. The comparative experimental results show that the operating cost is reduced by 15%.

[0356] 3. The combination of load prediction and rolling time domain optimization in the present invention significantly improves the system's ability to cope with load fluctuations. By predicting load changes through LSTM and using MPC for rolling optimization scheduling, the system can quickly adjust when the load fluctuates, and the response speed is improved to within 5 minutes. Compared with the traditional static scheduling method, the present invention has a faster response speed and is more flexible, and can effectively cope with rapid changes in power, heat and cooling loads.

[0357] Embodiment 2:

[0358] Reference Figure 2 In a second embodiment of the present invention, the present invention provides an energy storage scheduling device for a combined cooling, heating and power system, comprising:

[0359] The CCHP equipment modeling module is used to model the power generation equipment, heating equipment and refrigeration equipment in the CCHP system. The model covers the output power, energy efficiency characteristics and operating condition parameters of the equipment; it is used to establish the model of the energy storage device, including batteries, heat storage and cold storage devices, and describe the charging and discharging efficiency, capacity and response speed characteristics of the energy storage device;

[0360] The load forecasting module is used to forecast the cooling, heating and electricity loads in the future period, providing a basis for the formulation of dispatching strategies. The forecast is based on historical load data and environmental factors.

[0361] The objective function module is used to determine the scheduling objective function according to the operation requirements and strategies. The objective function contains multiple optimization objectives, including economy, energy efficiency and environmental impact;

[0362] The energy storage optimization scheduling module is used to optimize the charging and discharging of the energy storage device based on the objective function. The optimization algorithm is used to achieve the objective function requirements while satisfying the constraints.

[0363] Real-time dispatch control module, used to dispatch and adjust charging and discharging and cooling and heating production in real time based on real-time load and energy storage status to adapt to dynamically changing load demands;

[0364] The rolling horizon optimization module is used to continuously optimize the control strategy at each time step.

[0365] The energy storage scheduling device for the combined heat and power system provided in the embodiment of the present invention and the energy storage scheduling method for the combined heat and power system provided in Example 1 are based on the same technical concept, and can produce the beneficial effects described in Example 1. For the contents not described in detail in this embodiment, please refer to Example 1.

[0366] Embodiment three:

[0367] A terminal provided by an embodiment of the present invention includes a processor and a storage medium;

[0368] The storage medium is used to store instructions;

[0369] The processor is used to operate according to the instruction to execute the steps of the method according to any one of the embodiments.

[0370] Embodiment 4:

[0371] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of any method described in Embodiment 1 are implemented.

[0372] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0373] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0374] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0375] The computer program instructions for performing the operation of the present invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed entirely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be personalized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present invention.

[0376] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for energy storage scheduling of a combined cooling, heating and power system, characterized in that: The following steps are involved: S1. Collect and establish the physical models and data of each device in the cogeneration system; S2, load forecasting by using long short-term memory neural network LSTM to predict the demand for electricity, heating and cooling in future periods; S3. Based on the load forecast results, a scheduling optimization objective function is established, where the objective function defines the optimization direction and evaluation criteria, and the objective function includes minimizing operating costs, maximizing energy efficiency, and minimizing carbon emissions; S4. Use deep reinforcement learning algorithms to schedule the charging and discharging process of the energy storage device to meet the upper and lower limits of the equipment output, the limitations of the energy storage capacity, and the constraints of power balance; S5. During operation, model predictive control is used to make real-time scheduling adjustments based on real-time load and energy storage device status; S6. Use rolling horizon optimization to continuously optimize the scheduling strategy at each time step.

2. The energy storage scheduling method for a combined cooling, heating and power system according to claim 1, characterized in that: In the step S1, the operating data of various equipment in the cogeneration system are first collected, including the power output of the power generation equipment, the output power of the heating equipment and the refrigeration equipment, and the charging and discharging status of the battery, the heat storage device and the cold storage device. Then, the temperature and humidity of the environment are collected, and finally, the physical model of each device is established.

3. The energy storage scheduling method for a combined cooling, heating and power system according to claim 1, characterized in that: The step S2 includes: S201, preprocess the historical load data and environmental data, including normalization, standardizing the data to the [0,1] interval, S202, construct and train an LSTM model, with historical load data and environmental data as input and power, heating and cooling demand for multiple time steps in the future as output. The training goal is to minimize the mean square error (MSE), and the Adam optimization algorithm is used to update the model parameters. S203: Based on the real-time data input into the trained LSTM model, the power, heat and cooling demand in the future time period is predicted to generate the predicted load value L elec ,L heat ,L cool .

4. The energy storage scheduling method for a combined cooling, heating and power system according to claim 1, characterized in that: In the S3 step, the running cost is minimized: Among them, J1 is the total operating cost of each device, T is the time range of optimization, including T time steps, C f is the fuel cost of each device, F g is the fuel consumption, C grid is the electricity purchase cost, P grid，t is the power purchased from the grid at time step t, C storage is the operating cost of the energy storage device, P storage,t是 The charge and discharge power of the energy storage device at time step t, where positive values ​​represent discharge and negative values ​​represent charge; Maximize energy efficiency: Among them, J2 is the energy utilization efficiency, P out，t is the power output of the system at time step t, Q out，t is the heat output of the system at time step t, T is the time range of optimization, including T time steps, F g is the fuel consumption, which indicates the amount of fuel used by the system in time step t, P grid,t is the power purchased from the grid at time step t; Minimize Carbon Emissions: Among them, J3 is the carbon emission target, the goal is to minimize it, e g is the carbon emission factor of the fuel, which indicates the carbon emission generated per unit of fuel consumption, T is the optimization time range, including T time steps, F g is the fuel consumption, which represents the amount of fuel used by the system in time step t; The final comprehensive objective function is: J min =α1J1-α2J2+α3J3 J min : Minimize the comprehensive objective function, which represents the weighted sum of different objectives. J1 is the operating cost objective, J2 is the energy efficiency objective, and J3 is the carbon emission objective. α1 is the weight coefficient used to balance the importance of operating costs, α2 is the weight coefficient used to balance the importance of energy efficiency, and α3 is the weight coefficient used to balance the importance of carbon emissions. Constraint setting: including upper and lower limits of equipment output, energy storage capacity restrictions, and power balance constraints: P min，i ≤P i，t ≤P max，i 0≤E storage,t ≤C storage P gen,t +P storage,t =L elec,t +P grid,t P min,i ,P max,i are the minimum and maximum power output limits of device i; P i,t is the power output of device i at time step t E storage,t is the energy state of the energy storage device at time step t; C storage is the maximum capacity of the energy storage device; P gen,t is the electrical power output by the power generation equipment at time step t; P storage,t is the power output of the energy storage device at time step t. If the value is positive, it means that the energy storage device is discharging to supply the load; if the value is negative, it means that the energy storage device is charging; L elec,t is the power load demand at time step t, which indicates the amount of electricity required by the user at this point in time; P grid,t is the power purchased from the external grid at time step t.

5. The energy storage scheduling method for a combined cooling, heating and power system according to claim 1, characterized in that: In the step S4, the specific process includes: S401, State space and action space definition State space s t :The state space defines the state of the cogeneration system at a certain moment, which is used to describe the complete operation of the cogeneration system. The state contains the following variables: Energy storage state E storage,t : Indicates the current energy level of the energy storage device; Load Forecast elec,t ,L heat,t ,L cool,t : Future load demand prediction by LSTM; Current device status P gen,t , P grid,t : Including the output power of power generation equipment and external power purchase power; Environmental parameters T, H: current ambient temperature and humidity; Action Space a t : The action space defines the charging and discharging operations of the energy storage device: Energy storage charging and discharging power P storage,t : Charging is a positive value and discharging is a negative value, indicating the charging and discharging behavior of the energy storage device at time step t; S402. Reward function design R t =-(α1J1+α2J2+α3J3) Among them, R t is the reward value at time step t, which is used to guide the learning direction of the deep reinforcement learning algorithm, α1 is the weight coefficient used to balance the importance of operating cost, J1 is the operating cost target, α2 is the weight coefficient used to balance the importance of energy efficiency, J2 is the energy efficiency target, α3 is the weight coefficient used to balance the importance of carbon emissions, J3 is the carbon emission target; Minimization of operating costs J1: including fuel, electricity purchase and energy storage operating costs, increasing rewards by reducing operating costs; Maximizing energy efficiency J2: Increase the reward value by efficiently utilizing energy; Carbon emission minimization J3: Increase rewards by reducing carbon emissions; S403. Deep reinforcement learning training process Initialize the policy network and value network: Policy network: used to output in state s t The next action a t , i.e. how the energy storage device should be charged or discharged; Value network: used to evaluate the state s t Next, perform action a t The expected return Q(s) t ,a t ), the goal of the Critic network is to help the Actor network improve its strategy; Experience replay: When the system interacts with the environment, the data of each interaction is stored in the experience replay buffer: (s t ,a t ,R t ,s t+1 ) The network is trained by sampling small batches of data from the experience replay buffer to avoid correlation between data and improve the stability and efficiency of training; Target network soft update: In order to improve the stability of network training, the target network is used for updating. The parameters of the target network are obtained by soft updating of the policy network and the value network. The update rule is: θ′=τθ+(1-τ)θ′ Among them, θ′ represents the parameters of the target network, θ represents the parameters of the main network, and τ is 0.001, which is used to control the update speed. Policy improvement: Use the gradient ascent method to update the policy network, the goal is to increase the policy in state s t Next, perform action a t The return: is the gradient of the policy, that is, the update direction of the policy function; is the state s t Expected value; It is the gradient of the action value Q function with respect to action a, which indicates the rate of change of the action value function under state s and action a; s=s t , a=μ(s t ) refers to the action selected by the current policy network μ in state st; is the gradient of the policy function μ with respect to its parameter θμ; The value network update uses gradient descent to minimize the value network's loss function, which is the mean squared error between the actual return and the target Q value: Among them, L(θ Q ) is the loss function, which is used to measure the estimation error of the Q function; θ Q are the parameters of the action-value Q function; E(s t ,a t ,R t ,s t+1 ) is the expected value of a set of sample data, and the expectation is based on the sampling in the experience pool (s t ,a t ,R t ,s t+1 ) quaternion; Q(s t ,a t ∣θ Q ) is the estimated Q value of the state st and action at given by the current Q function; y t is the target Q value, used to train the Q function: y t =R t +γQ′(s t+1 ,μ′(s t+1 )|θ Q′ ) R t It’s an instant reward; γ is a discount factor, which indicates the importance of future rewards; Q′(s t+1 ,μ′(s t+1 )|θ Q′ ) is the next state s given by the target Q network t+1 The estimated value of is calculated using the target Q network and the target policy network; θ Q′ are the parameters of the target Q network; S404. Execution of energy storage optimization scheduling Action selection: At each time step t, use the policy network based on the current state s t Select action a t , which determines the charging and discharging power of the energy storage device at the current moment; Action a t Will directly affect the energy state E of the energy storage device storage,t+1 , thus affecting future energy supply and demand; Perform charge and discharge operations: According to the selected action at, the corresponding charging and discharging operations are performed, the energy state of the energy storage device is changed, and the future operation state is affected; Interact with the environment and give feedback: After executing an action, the system interacts with the environment and generates a new state s t+1 and reward R t , these new states will serve as input for the next decision, enter the experience replay buffer, and be used for further strategy improvement.

6. The energy storage scheduling method for a combined cooling, heating and power system according to claim 1, characterized in that: In the S5 step, the optimal strategy learned by deep reinforcement learning during the training process will be used as the initial plan for real-time scheduling control. During actual operation, model predictive control will optimize and adjust the strategy provided by deep reinforcement learning in real time based on the current state and load forecast to cope with dynamic changes in the environment.

7. The energy storage scheduling method for a combined cooling, heating and power system according to claim 1, characterized in that: The step S6 includes: S601, load forecast update: at each time step, use real-time data to input the LSTM model to obtain new future load forecast values, which will affect the next dispatch decision; S602, optimization solution update: Combine the updated load forecast and the current state, use the strategy trained by deep reinforcement learning to optimize the energy storage equipment and power generation equipment, and recalculate the optimal charging and discharging plan and power output at each time step; S603, control strategy execution: the optimal control strategy obtained by solving is executed at each time step, and refined and adjusted through model predictive control.

8. An energy storage dispatching device for a combined cooling, heating and power system, characterized in that: The energy storage scheduling method for the combined cooling, heating and power system according to any one of claims 1 to 7, the device comprising: The CCHP equipment modeling module is used to model the power generation equipment, heating equipment and refrigeration equipment in the CCHP system. The model covers the output power, energy efficiency characteristics and operating condition parameters of the equipment; it is used to establish the model of the energy storage device, including batteries, heat storage and cold storage devices, and describe the charging and discharging efficiency, capacity and response speed characteristics of the energy storage device; The load forecasting module is used to forecast the cooling, heating and electricity loads in the future period, providing a basis for the formulation of dispatching strategies. The forecast is based on historical load data and environmental factors. The objective function module is used to determine the scheduling objective function according to the operation requirements and strategies. The objective function contains multiple optimization objectives, including economy, energy efficiency and environmental impact; The energy storage optimization scheduling module is used to optimize the charging and discharging of the energy storage device based on the objective function. The optimization algorithm is used to achieve the objective function requirements while satisfying the constraints. Real-time dispatch control module, used to dispatch and adjust charging and discharging and cooling and heating production in real time based on real-time load and energy storage status to adapt to dynamically changing load demands; The rolling horizon optimization module is used to continuously optimize the control strategy at each time step.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the energy storage scheduling method for the combined heat and power system according to any one of claims 1 to 7 is implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the energy storage scheduling method for a combined heat and power system according to any one of claims 1 to 7 is implemented.

Images