Multi-energy-flow scheduling method for cooling, heating and power integrated energy system considering source load uncertainty

Abstract

The invention discloses a multi-energy-flow scheduling method for a cooling, heating and power integrated energy system with consideration of source load uncertainty. The method comprises the following steps: 1, establishing physical models of equipment of the cooling, heating and power integrated energy system; 2, establishing operation constraints and security constraints of the system, and determining schedulable variables of the system in the operation process; 3, integrating system operation characteristics and schedulable variables which relate to uncertainty, and establishing a Markov decision process MDP model; 4, improving a depth deterministic policy gradient (DDPG) algorithm; and 5, network training in a high-dimensional state action space environment is carried out based on the improved ME-DDPG algorithm, and efficient real-time optimization scheduling of the cooling, heating and power system is realized. The problem that multi-energy flow decoupling and scheduling optimization in a certain specific scene are insufficient in multi-energy flow optimization research in a source load uncertainty scene is solved. The ME-DDPG algorithm has relatively high scheduling implementation efficiency, and the adaptive capacity of the system to the source load uncertainty is remarkably enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of cooling, heating and power cogeneration system optimization, and in particular relates to a multi-energy flow scheduling method for a cooling, heating and power integrated energy system taking into account source and load uncertainty. Background Art

[0002] In the context of promoting the dual carbon goals, the development and utilization of renewable energy resources have received extensive attention, and their integrated application in energy systems has increased. However, the intermittent, volatile and uncertain characteristics of renewable energy restrict its stable energy supply and large-scale application. Integrated energy systems have become a hot topic of research because they have demonstrated good renewable energy absorption capacity by integrating traditional energy and renewable energy. Current research mainly focuses on combined heat and power (CHP) and combined cooling, heating and power (CCHP) systems, and reduces fuel consumption by coupling with renewable energy sources such as photovoltaics and wind power. In addition, by integrating energy storage devices, the peak-shaving capacity of the system is improved, thereby reducing fuel consumption to a certain extent, reducing system costs and enhancing economy. However, existing research focuses on performance analysis in specific scenarios. There is still a lack of research on multi-energy flow optimization in source-load uncertainty scenarios, especially the efficient and intelligent scheduling of wind-solar coupled cooling, heating and power integrated energy systems. As a problem-solving method suitable for complex dynamic environments, reinforcement learning (RL) has the core advantage of being able to make decisions and learn in unknown environments through interaction with the environment, experience accumulation and feedback. Deep reinforcement learning (DRL) uses neural networks to process high-dimensional and continuous state spaces, making it a powerful tool for solving complex real-world problems. Through deep neural networks for function approximation, DRL effectively alleviates the computational difficulties of processing complex environments and large-scale state spaces, overcoming the "curse of dimensionality" challenge faced by traditional algorithms. Summary of the invention

[0003] Purpose of the invention: In view of the problem that the existing technology usually only involves the decoupling and scheduling optimization of multiple energy flows in a certain specific scenario, and the research on the optimization of multiple energy flows in scenarios with source and load uncertainty is still insufficient, the purpose of the present invention is to provide a multiple energy flow scheduling method for a cold, heat and power integrated energy system taking into account source and load uncertainty.

[0004] Technical solution: A multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty of the present invention comprises the following steps:

[0005] Step 1: Based on the characteristics of each device in the system, establish the physical model of each device in the cooling, heating and power integrated energy system;

[0006] Step 2: Establish the operation constraints and safety constraints of the integrated cooling, heating and power energy system, consider the energy flow conversion and circulation relationship between the various devices in the system, and determine the dispatchable variables of the system during operation;

[0007] Step 3: Integrate the system operation characteristics and dispatchable variables involving uncertainty, establish a Markov decision process MDP model, and transform the scheduling optimization problem into a sequential decision problem;

[0008] Step 4: Introduce a multi-exploration mechanism and design a time-sharing reward function for the time-sharing electricity price problem to improve the deep deterministic policy gradient DDPG algorithm and obtain the improved deep deterministic policy gradient ME-DDPG algorithm;

[0009] Step 5: Based on the improved ME-DDPG algorithm, network training is performed in a high-dimensional state-action space environment to achieve efficient real-time optimization scheduling of the cold, heat, electricity and hydrogen system.

[0010] Furthermore, in step 1, the integrated cooling, heating and electricity energy system includes a fuel cell, an electrolyzer, a heat pump, a refrigeration device, a renewable energy power generation device, an energy storage device, a power grid and cooling, heating and electricity loads.

[0011] Furthermore, the fuel cell is a proton exchange membrane fuel cell PEMFC, and its physical property model is:

[0012]

[0013] Among them, N FC is the hydrogen consumption rate of the fuel cell, n cell is the number of cells contained in the fuel cell, I is the current of the fuel cell, and F is the Faraday constant; current density is usually used to characterize the intensity and rate of chemical reactions in PEMFC. The current density equation is defined as follows:

[0014]

[0015] Where j represents the current density, A cell is the effective area of ​​a unit cell in a fuel cell;

[0016] The electrolyzer, whose internal battery is a proton exchange membrane electrolysis battery, has a physical property model as follows:

[0017]

[0018] Among them, η F is the Faraday efficiency, and I EH Respectively represent the number of cells and current in the electrolytic cell, N EH is the hydrogen production rate of water electrolysis;

[0019] The heat pump is an air source heat pump, and its physical characteristic model is:

[0020] ΔT HP=T DHW -T e

[0021]

[0022]

[0023] Among them, T DHW is the hot water temperature in the heat pump, T e is the current ambient temperature, h1, h2, h3 are all empirical fitting coefficients, is the electrical power consumed by the heat pump, is the thermal power generated by the heat pump;

[0024] The refrigeration equipment includes an electric refrigerator and an absorption refrigerator, and its physical characteristic model is:

[0025] The physical characteristic model of the electric refrigerator is:

[0026]

[0027] Among them, η ECh is the cooling efficiency of the electric refrigerator, and Respectively represent the electrical power consumed and the cooling power generated by the electric refrigerator;

[0028] The physical characteristic model of the absorption refrigerator is:

[0029]

[0030] Among them, η ACh is the refrigeration efficiency of the absorption chiller, and are the heat power consumed and the cold power generated by the absorption chiller respectively;

[0031] The renewable energy power generation equipment includes photovoltaic equipment and wind turbine generator sets, and its physical characteristic model is:

[0032] The physical characteristic model of the photovoltaic device is:

[0033]

[0034] Among them, η PV is the photovoltaic efficiency, η inv is the inverter efficiency, I 0 is the solar irradiance, expressed in watts per square meter (W / m 2 ), A PV is the effective lighting area;

[0035] The physical characteristic model of the wind turbine generator set is:

[0036]

[0037] Among them, v t , v in , v out and v r They are actual wind speed, startup wind speed, shutdown wind speed and rated wind speed, respectively. wt is the installed capacity of the wind turbine generator set, is the wind power generation power;

[0038] The energy storage device includes a heat storage tank and a hydrogen storage tank, and its physical characteristic model is:

[0039] The physical characteristic model of the heat storage tank is:

[0040]

[0041] Among them, HSD t is the energy storage state of the heat storage tank at time t, HSD t+1 is the energy storage state after updating at time t+1, H c is the maximum capacity of the heat storage tank, ΔT is the time step, Indicates the working status of the heat storage tank. A positive number indicates the heat entering the heat storage tank, and a negative number indicates the heat released by the heat storage tank.

[0042] The physical property model of the hydrogen storage tank is:

[0043]

[0044]

[0045] Among them, TP t and are the hydrogen storage state and hydrogen flow rate of the hydrogen storage tank at time t, V H , T H , M H2 and R are volume, temperature, molar mass of hydrogen and gas constant, respectively. t is the compression coefficient of hydrogen, T0, a i , b i and c i All are constants, and T0=100K.

[0046] Furthermore, step 2 is specifically as follows:

[0047] Establish the following safe operating constraints:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] HSD min ≤HSD t ≤HSD max

[0057] TP min ≤TP t ≤TP max

[0058] in, and are the upper and lower limits of the fuel cell output power, and are the upper and lower limits of the electrolytic power of the electrolytic cell, and are the upper and lower limits of the heat pump power consumption, and are the upper and lower limits of the power consumption of the electric refrigerator, and are the upper and lower limits of heat storage tank charging and discharging, respectively. and are the upper and lower limits of the heat absorbed by the absorption refrigerator, and They are the upper and lower limits of hydrogen storage tank charging and discharging, HSD max and HSD min are the maximum and minimum values ​​of HSD in the heat storage tank, TP max and TP min are the maximum and minimum values ​​of TP in the energy storage state of the hydrogen storage tank, respectively;

[0059] The following power balance constraints are established:

[0060]

[0061]

[0062]

[0063]

[0064] in, and are the electrical, heating and cooling loads of the system at time t, The positive and negative values ​​of represent the purchase of electricity from the grid and the sale of electricity to the grid. The positive and negative values ​​respectively indicate the release of hydrogen from the hydrogen storage tank and the filling of hydrogen into the hydrogen storage tank;

[0065] The following objective function is established:

[0066]

[0067]

[0068] Among them, k represents different types of equipment, OM k is the unit operation and maintenance cost of device k, P k,t is the output power of device k at time t; and They represent the power purchased and sold from the power grid, p in,t and p out,t are the unit prices of electricity purchase and electricity sales, respectively, λ1 and λ2 are parameters for adjusting the size of the two parts of the cost, HSD T and HSD0 are the state values ​​of the heat storage tank at the final and initial moments, respectively. T and TP0 are the status values ​​of the hydrogen storage tank at the final and initial moments, respectively.

[0069] Furthermore, step 3 is specifically: a sequential decision problem based on a Markov decision process, including: establishing the following state space:

[0070]

[0071] Among them, the state space s t Including time, electrical load, heat load, cooling load, wind power generation power, photovoltaic power generation power, heat storage status of the thermal storage tank and hydrogen storage status of the hydrogen storage tank;

[0072] Create the following action space:

[0073]

[0074] Among them, the action space a t Including the power generation of fuel cells, the power consumption of heat pumps, the power consumption of electric refrigerators and the power consumption of water electrolysis;

[0075] Create the following reward function:

[0076]

[0077]

[0078]

[0079] Among them, the reward function r t Including system operation and maintenance costs C OM,t , the cost of electricity purchase and sale from the power grid C net,t , supply and demand imbalance penalty C out,t , thermal imbalance Pressure imbalance Cold imbalance And the energy storage equipment status cost C T , water electrolysis power μ i (i=1,2,3) and λ i (i=1,...,4) are all parameters that need to be adjusted.

[0080] Furthermore, in step 4, the improved deep deterministic policy gradient ME-DDPG algorithm includes the following two parts:

[0081] Improvement 1: Multiple exploration mechanisms; When implementing the deep deterministic policy gradient (DDPG) algorithm for training, exploration noise is generated through the Ornstein-Uhlenbeck noise generation mechanism; then, the opposite of the noise is superimposed on the decision action output by the policy network, while retaining the original action decision, thereby forming three different behavior strategies; in the process of interacting with the environment, the three strategies each generate independent experience trajectories; then, the experience trajectory with the largest reward value is selected and included in the experience replay pool for subsequent learning and training iterations;

[0082] Improvement 2: Time-sharing reward function; In order to solve the challenges of time-sharing electricity prices encountered during system operation, the reward mechanism is dynamically adjusted to increase the reward for water electrolysis to produce hydrogen during valley electricity prices to encourage hydrogen production during this period; and during peak electricity prices, the reward for water electrolysis to produce hydrogen is reduced, and even penalties can be imposed to suppress hydrogen production during periods with higher electricity prices.

[0083] The above two improvements not only improve learning efficiency, enhance the generalization ability of the algorithm in complex environments, and avoid the problem of premature convergence to local optimal solutions, but also encourage the strategy network to reasonably arrange the activities of water electrolysis to produce hydrogen according to electricity price fluctuations, thereby optimizing energy utilization efficiency, reducing costs, and improving the economy and sustainability of the entire system.

[0084] Furthermore, step 5 is specifically: network training in a high-dimensional state-action space environment based on the improved ME-DDPG algorithm, including:

[0085] Initialize the policy network P, evaluation network Q1, Q2 and their corresponding replica networks, namely the target policy network P', target evaluation network Q1', Q'2; the construction of the policy network is to receive the feature vector representing the operating state of the system as input, and map it to the schedulable decision variable through the network structure; at the same time, the architecture of the evaluation network is designed to process the joint input of state and action, and its output is the corresponding reward signal;

[0086] A random source load fluctuation is introduced, and the fluctuation amplitude is limited to no more than 30% of the baseline scenario load to randomly generate a set of training scenarios. In each time step of each training round, two improved strategies, multi-exploration mechanism and time-sharing reward function, are adopted to perform environmental exploration and training experience accumulation. The accumulated training experience consists of a four-tuple, including state, action, reward and next state, and its mathematical expression is as follows: (s t ,a t ,r t ,s t+1 );

[0087] After the neural network training process is completed, the policy network model is saved to be applied to the efficient and intelligent scheduling of the integrated cooling, heating and power energy system. In the real-time operation scenario, the policy network can quickly make decisions based on the immediate state of the system and output an optimized scheduling strategy, thereby improving energy utilization efficiency and minimizing the operation and maintenance costs of the system.

[0088] The present invention also discloses a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0089] The present invention also discloses a computer-readable storage medium on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the method of the present invention are implemented.

[0090] The present invention also discloses a computer program product, comprising a computer program / instruction, which implements the steps of the method of the present invention when executed by a processor.

[0091] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0092] Flexibility: Compared to rule-based algorithms, the present invention does not rely on a predefined set of rules, thereby ensuring stable operation of the system in the face of unforeseen situations.

[0093] Computational efficiency: Compared with traditional learning-based algorithms, although the computational workload and computation time are larger when dealing with complex systems, the present invention avoids the problem of finding suboptimal solutions only for specific scenarios and shows better performance in uncertain environments.

[0094] Dimensional processing: Traditional reinforcement learning algorithms are prone to dimensionality disasters when dealing with high-dimensional state and action spaces, resulting in a surge in storage space requirements. This invention uses deep reinforcement learning technology and deep neural networks for function approximation, effectively alleviating the difficulty of dealing with complex environments and large-scale state spaces.

[0095] Based on the deep reinforcement learning algorithm DDPG, this paper makes the following two improvements:

[0096] Multi-exploration mechanism: Exploration noise is generated by adopting the Ornstein-Uhlenbeck noise generation mechanism. The inverse of this noise is combined with the action output by the policy network while retaining the original action to form three different behavioral strategies. Each strategy generates an independent experience trajectory when interacting with the environment. On this basis, the experience trajectory with the highest reward value is selected and stored in the experience replay pool for subsequent learning and training iterations.

[0097] Time-sharing reward function: The present invention designs a time-sharing reward function to adapt to the periodic changes in the electricity market. During the low electricity price period, the reward for the water electrolysis hydrogen production process is increased, while during the peak electricity price period, the reward for the water electrolysis hydrogen production process is reduced.

[0098] Thanks to the above two improvement measures, the algorithm of the present invention has achieved significant improvements in convergence speed and convergence effect, thereby providing a more solid algorithmic foundation for efficient and intelligent scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 This is a schematic diagram of the structure of a cold, heat and electricity integrated energy system according to an embodiment of the present invention;

[0100] Figure 2 This is a flow chart of a cold, heat and electricity integrated energy system according to an embodiment of the present invention.

[0101] Figure 3 Meteorological data of a typical day scene in an embodiment of the present invention;

[0102] Figure 4 This is the load data of a typical day scenario in an embodiment of the present invention.

[0103] Figure 5 These are the training curves of the DDPG algorithm and the ME-DDPG algorithm according to the embodiments of the present invention.

[0104] Figure 6This is a diagram of energy storage device status changes in a typical day scenario according to an embodiment of the present invention. DETAILED DESCRIPTION

[0105] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0106] In this embodiment, the specification parameters and constraints of each device included in the system are shown in Table 1:

[0107] Table 1 Main parameters of each equipment

[0108]

[0109]

[0110] The main training parameters of the DDPG algorithm are listed in Table 2. In view of the order of magnitude difference between the system operation and maintenance cost and the grid purchase and sale cost, the grid purchase and sale cost coefficient is appropriately adjusted and set to a smaller value to ensure that the operation and maintenance cost is at the same order of magnitude as the grid purchase and sale cost, thereby avoiding the adverse effect of the grid purchase and sale cost on the algorithm training effect. Before each training round, based on the load data under typical daily scenarios (shown in Figure 3 and Figure 4 ), a random source load fluctuation of no more than 30% of the current load is introduced to generate new training scenarios. In order to verify the effectiveness of the algorithm improvement measures, the training scenario load, action noise, and initial weight data of the policy network and evaluation network of each training round are retained during the execution of the DDPG algorithm. These data will be used as the training load, exploration noise, and initialization parameters of the policy and evaluation networks of the improved DDPG (ME-DDPG). The training process of the two algorithms is Figure 5 The training curve in is shown, where the "mean" subscript represents the average value of the training reward every 5 steps. After the training is completed, the policy network is saved and simulated for optimization scheduling in a typical day scenario. Based on this, the comprehensive scheduling cost of the system within 24 hours (including system operation cost and grid power purchase and sales cost), the final state of the heat storage tank and hydrogen storage tank, and the execution time of a single-step scheduling are obtained. In addition, Figure 6 The changes in the energy storage status of each energy storage device under a typical day scenario are shown.

[0111] according to Figure 5 The results shown show that, while keeping the training scenario, network initialization parameters, and exploration noise consistent, the ME-DDPG algorithm converges in about 170 training rounds, which is about 11.76% faster than the standard DDPG algorithm, which converges in 190 training rounds. In addition, Figure 5The sub-figure in shows the average reward values ​​of the two algorithms during the 200th to 240th rounds. Specifically, the average training reward of the ME-DDPG algorithm after convergence is -14.98, while the average training reward of the DDPG algorithm after convergence is -17.47. Therefore, the performance of the ME-DDPG algorithm is about 14.25% better than that of the DDPG algorithm. The statistical analysis of the training results shows that the proposed ME-DDPG algorithm outperforms the classic DDPG algorithm in terms of convergence speed and final convergence reward. This result confirms the effectiveness of the two improvements to the DDPG algorithm in this study.

[0112] Table 3 shows the dispatch results in a typical day scenario, from which the following conclusions can be drawn: In the comparison of comprehensive dispatch costs, the ME-DDPG algorithm achieved a 3.23% reduction compared to the DDPG algorithm. In terms of the energy storage state of the energy storage device at the end of the dispatch, both algorithms can effectively restore the energy storage device to near the initial state. However, thanks to its effective improvement measures, the ME-DDPG algorithm makes the energy storage state closer to the initial moment at the end of the dispatch. In the evaluation of the single-step dispatch time, the single-step dispatch time of the ME-DDPG algorithm is shortened to 0.24 seconds, which shows that the ME-DDPG algorithm has a high dispatch implementation efficiency and significantly enhances the system's adaptability to source and load uncertainties. Therefore, the proposed ME-DDPG algorithm not only ensures the dispatch efficiency, but also improves the dispatch performance of the system when facing source and load fluctuations.

[0113] Table 2 Main training parameters of ME-DDPG

[0114]

[0115] Table 3 Comparison of scheduling results for typical day scenarios

[0116]

Claims

1. A multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty, characterized in that: The steps include: Step 1: Based on the characteristics of each device in the system, establish the physical model of each device in the cooling, heating and power integrated energy system; Step 2: Establish the operation constraints and safety constraints of the integrated cooling, heating and power energy system, consider the energy flow conversion and circulation relationship between the various devices in the system, and determine the dispatchable variables of the system during operation; Step 3: Integrate the system operation characteristics and dispatchable variables involving uncertainty, establish a Markov decision process MDP model, and transform the scheduling optimization problem into a sequential decision problem; Step 4: Introduce a multi-exploration mechanism and design a time-sharing reward function for the time-sharing electricity price problem to improve the deep deterministic policy gradient DDPG algorithm and obtain the improved deep deterministic policy gradient ME-DDPG algorithm; Step 5: Based on the improved ME-DDPG algorithm, network training is performed in a high-dimensional state-action space environment to achieve efficient real-time optimization scheduling of the cooling, heating and power systems.

2. According to claim 1, a multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty is characterized in that: In step 1, the integrated cooling, heating and electricity energy system includes a fuel cell, an electrolyzer, a heat pump, a refrigeration device, a renewable energy power generation device, an energy storage device, a power grid and cooling, heating and electricity loads.

3. According to claim 2, a multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty is characterized in that: The fuel cell is a proton exchange membrane fuel cell PEMFC, and its physical characteristic model is: Among them, N FC is the hydrogen consumption rate of the fuel cell, n cell is the number of cells contained in the fuel cell, I is the current of the fuel cell, and F is the Faraday constant; current density is usually used to characterize the intensity and rate of chemical reactions in PEMFC. The current density equation is defined as follows: Where j represents the current density, A cell is the effective area of ​​a unit cell in a fuel cell; The electrolyzer, whose internal battery is a proton exchange membrane electrolysis battery, has a physical property model as follows: Among them, η F is the Faraday efficiency, and I EH Respectively represent the number of cells and current in the electrolytic cell, N EH is the hydrogen production rate of water electrolysis; The heat pump is an air source heat pump, and its physical characteristic model is: ΔT HP =T DHW -T e Among them, T DHW is the hot water temperature in the heat pump, T e is the current ambient temperature, h1, h2, h3 are all empirical fitting coefficients, is the electrical power consumed by the heat pump, is the thermal power generated by the heat pump; The refrigeration equipment includes an electric refrigerator and an absorption refrigerator, and its physical characteristic model is: The physical characteristic model of the electric refrigerator is: Among them, η ECh is the cooling efficiency of the electric refrigerator, and They represent the electrical power consumed and the cooling power generated by the electric refrigerator respectively; The physical characteristic model of the absorption refrigerator is: Among them, η ACh is the refrigeration efficiency of the absorption chiller, and are the heat power consumed and the cold power generated by the absorption chiller respectively; The renewable energy power generation equipment includes photovoltaic equipment and wind turbine generator sets, and its physical characteristic model is: The physical characteristic model of the photovoltaic device is: Among them, η PV is the photovoltaic efficiency, η inv is the inverter efficiency, I 0 is the solar irradiance, in watts per square meter, A PV is the effective lighting area; The physical characteristic model of the wind turbine generator set is: Among them, v t , v in , v out and v r They are actual wind speed, start-up wind speed, shutdown wind speed and rated wind speed, respectively. wt is the installed capacity of the wind turbine generator set, is the wind power generation power; The energy storage device includes a heat storage tank and a hydrogen storage tank, and its physical characteristic model is: The physical characteristic model of the heat storage tank is: Among them, HSD t is the energy storage state of the heat storage tank at time t, HSD t+1 is the energy storage state after updating at time t+1, H c is the maximum capacity of the heat storage tank, ΔT is the time step, Indicates the working status of the heat storage tank. A positive number indicates the heat entering the heat storage tank, and a negative number indicates the heat released by the heat storage tank. The physical property model of the hydrogen storage tank is: Among them, TP t and are the hydrogen storage state and hydrogen flow rate of the hydrogen storage tank at time t, V H , T H , M H2 and R are volume, temperature, molar mass of hydrogen and gas constant, respectively. t is the compression coefficient of hydrogen, T0, a i , b i and c i All are constants, and T0=100K.

4. According to claim 1, a multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty is characterized in that: Step 2 is as follows: Establish the following safe operating constraints: HSD min ≤HSD t ≤HSD max TP min ≤TP t ≤TP max in, and are the upper and lower limits of the fuel cell output power, and are the upper and lower limits of the electrolytic power of the electrolytic cell, and are the upper and lower limits of the heat pump power consumption, and are the upper and lower limits of the power consumption of the electric refrigerator, and are the upper and lower limits of heat storage tank charging and discharging, respectively. and are the upper and lower limits of the heat absorbed by the absorption refrigerator, and They are the upper and lower limits of hydrogen storage tank charging and discharging, HSD max and HSD min are the maximum and minimum values ​​of HSD in the heat storage tank, TP max and TP min are the maximum and minimum values ​​of TP in the energy storage state of the hydrogen storage tank, respectively; The following power balance constraints are established: in, and are the electrical, heating and cooling loads of the system at time t, The positive and negative values ​​of represent the purchase of electricity from the grid and the sale of electricity to the grid. The positive and negative of represent the release of hydrogen from the hydrogen storage tank and the filling of hydrogen into the hydrogen storage tank respectively; the following objective function is established: Among them, k represents different types of equipment, OM k is the unit operation and maintenance cost of device k, P k,t is the output power of device k at time t; and They represent the power purchased and sold from the power grid, p in,t and p out,t are the unit prices of electricity purchase and electricity sales, respectively, λ1 and λ2 are parameters for adjusting the size of the two parts of the cost, HSD T and HSD0 are the state values ​​of the heat storage tank at the final and initial moments, respectively. T and TP0 are the status values ​​of the hydrogen storage tank at the final and initial moments, respectively.

5. According to claim 1, a multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty is characterized in that: Step 3 is specifically: a sequential decision problem based on a Markov decision process, including: establishing the following state space: Among them, the state space s t Including time, electrical load, heat load, cooling load, wind power generation power, photovoltaic power generation power, heat storage status of the thermal storage tank and hydrogen storage status of the hydrogen storage tank; Create the following action space: Among them, the action space a t Including the power generation of fuel cells, the power consumption of heat pumps, the power consumption of electric refrigerators and the power consumption of water electrolysis; Create the following reward function: Among them, the reward function r t Including system operation and maintenance costs C OM,t , the cost of electricity purchase and sale from the power grid C net,t , supply and demand imbalance penalty C out,t , thermal imbalance Pressure imbalance Cold imbalance And the energy storage equipment status cost C T , water electrolysis power μ i (i=1,2,3) and λ i (i=1,...,4) are all parameters that need to be adjusted.

6. According to claim 1, a multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty is characterized in that: In step 4, the improved deep deterministic policy gradient ME-DDPG algorithm includes the following two parts: Improvement 1: Multiple exploration mechanisms: When implementing the deep deterministic policy gradient (DDPG) algorithm for training, exploration noise is generated through the Ornstein-Uhlenbeck noise generation mechanism; then, the opposite of the noise is superimposed on the decision action output by the policy network, while retaining the original action decision, thus forming three different behavior strategies; In the process of interacting with the environment, the three strategies each generate independent experience trajectories. Then, the experience trajectory with the largest reward value is selected and included in the experience replay pool for subsequent learning and training iterations. Improvement 2: Time-sharing reward function; In order to solve the challenges of time-sharing electricity prices encountered during system operation, the reward mechanism is dynamically adjusted to increase the reward for water electrolysis to produce hydrogen during valley electricity prices to encourage hydrogen production during this period; and during peak electricity prices, the reward for water electrolysis to produce hydrogen is reduced, and penalties may be imposed to inhibit hydrogen production during periods with higher electricity prices.

7. According to claim 1, a multi-energy flow scheduling method for a cold, heat and power integrated energy system considering source and load uncertainty is characterized in that: Step 5 is as follows: network training in a high-dimensional state-action space environment is performed based on the improved ME-DDPG algorithm, including: Initialize the policy network P, evaluation network Q1, Q2 and their corresponding replica networks, namely the target policy network P', target evaluation network Q1', Q'2; the construction of the policy network is to receive the feature vector representing the operating state of the system as input, and map it to the schedulable decision variable through the network structure; at the same time, the architecture of the evaluation network is designed to process the joint input of state and action, and its output is the corresponding reward signal; A random source load fluctuation is introduced, and the fluctuation amplitude is limited to no more than 30% of the baseline scenario load to randomly generate a set of training scenarios. In each time step of each training round, two improved strategies, multi-exploration mechanism and time-sharing reward function, are adopted to perform environmental exploration and training experience accumulation. The accumulated training experience consists of a four-tuple, including state, action, reward and next state, and its mathematical expression is as follows: (s t ,a t ,r t ,s t+1 ); After the neural network training process is completed, the strategy network model is saved to be applied to the efficient and intelligent scheduling of the integrated cooling, heating and power energy system.

8. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 1.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images