A method and apparatus for establishing a data-driven model

Abstract

The application discloses a kind of data-driven model establishing method and device, the method includes: according to the relevant parameter data of multi-energy park, the minimum cost objective function of multi-energy park and corresponding constraint condition are established;According to the deep reinforcement learning algorithm, the minimum cost objective function of multi-energy park is solved, and running state data are generated;According to the mapping relationship between the running state data and load data after the establishment of Gaussian process regression algorithm, the running state data is converted into new load data according to the mapping relationship;According to new load data, update load database, and establish the data-driven model of multi-energy park according to load database. The precision of the data-driven model of the multi-energy complementary park generated by the embodiment of the application can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for establishing a data-driven model. Background Technology

[0002] Multi-energy parks contain subsystems with various energy sources, and the energy flow of each subsystem is converted in different forms through coupling components. Currently, there is limited research on generating operational data for each system that meets planning requirements based on the coordination of multiple energy systems and by using data such as land use, area, power of each system, and installed capacity. Furthermore, because multi-energy complementary parks encompass a diverse range of equipment, compared to traditional energy systems, the operation of each subsystem in a multi-energy park depends not only on its own equipment but also on the status of other subsystems, resulting in complex operational conditions.

[0003] However, in the existing technology, due to the insufficient amount of multi-energy park operation scenario data and load data, it is impossible to conduct energy efficiency assessment or comprehensive scenario generation of the park based on a small amount of load data, resulting in low accuracy of the data-driven model of the generated multi-energy complementary park. Summary of the Invention

[0004] This invention provides a method and apparatus for establishing a data-driven model, which can effectively improve the accuracy of the generated data-driven model for a multi-energy complementary park.

[0005] The first aspect of this application provides a method for establishing a data-driven model, including:

[0006] Based on the relevant parameter data of the multi-functional park, establish the objective function for minimizing the cost of the multi-functional park and the corresponding constraints.

[0007] The objective function for minimizing the cost of the multi-functional park is solved using a deep reinforcement learning algorithm, and operational status data is generated.

[0008] Based on the Gaussian process regression algorithm, after establishing the mapping relationship between operating status data and load data, the operating status data is transformed into new load data according to the mapping relationship;

[0009] The load database is updated based on the new load data, and a data-driven model for the multi-energy park is established based on the load database.

[0010] In one possible implementation of the first aspect, the objective function for minimizing the cost of the multi-functional park is specifically as follows:

[0011] C T =min(C e +C m );

[0012] Among them, CT C represents the total cost of operating a multi-functional park. e Represents operating costs, C m This represents maintenance costs.

[0013] In one possible implementation of the first aspect, the corresponding constraints include:

[0014] Constraints on energy storage batteries, constraints on power-to-gas conversion equipment, constraints on gas turbines, and constraints on the supply and demand balance of natural gas and power systems.

[0015] In one possible implementation of the first aspect, the operational status data is generated by solving the objective function of minimizing the cost of the multi-functional park using a deep reinforcement learning algorithm, specifically as follows:

[0016] Based on the objective function of minimizing the cost of the multi-functional park and the corresponding constraints, a model for generating operational status data is generated.

[0017] The runtime data generation model is transferred to a Markov game, and a deep reinforcement learning algorithm is used to solve the runtime data generation model to calculate the runtime data.

[0018] In one possible implementation of the first aspect, a mapping relationship between operating status data and load data is established based on the Gaussian process regression algorithm, specifically as follows:

[0019] A machine learning model is trained and established based on the Gaussian process regression algorithm; the machine learning model is used to establish the mapping relationship between operating status data and load data.

[0020] A second aspect of this application provides a data-driven model building apparatus, comprising: a first building module, a second building module, a calculation module, and a third building module;

[0021] The first module is used to establish the objective function for minimizing the cost of the multi-energy park and the corresponding constraints based on the relevant parameter data of the multi-energy park.

[0022] The second module is used to solve the objective function of minimizing the cost of the multi-functional park based on the deep reinforcement learning algorithm, and generate operational status data.

[0023] The calculation module is used to establish a mapping relationship between operating status data and load data based on the Gaussian process regression algorithm, and then convert the operating status data into new load data according to the mapping relationship.

[0024] The third module is used to update the load database based on the new load data and to build a data-driven model for the multi-energy park based on the load database.

[0025] In one possible implementation of the second aspect, the objective function for minimizing the cost of the multi-functional park is specifically as follows:

[0026] C T =min(C e +C m );

[0027] Among them, C T C represents the total cost of operating a multi-functional park. e Represents operating costs, C m This represents maintenance costs.

[0028] In one possible implementation of the second aspect, the corresponding constraints include:

[0029] Constraints on energy storage batteries, constraints on power-to-gas conversion equipment, constraints on gas turbines, and constraints on the supply and demand balance of natural gas and power systems.

[0030] In one possible implementation of the second aspect, the operational status data is generated by solving the objective function of minimizing the cost of the multi-functional park using a deep reinforcement learning algorithm. Specifically:

[0031] Based on the objective function of minimizing the cost of the multi-functional park and the corresponding constraints, a model for generating operational status data is generated.

[0032] The runtime data generation model is transferred to a Markov game, and a deep reinforcement learning algorithm is used to solve the runtime data generation model to calculate the runtime data.

[0033] In one possible implementation of the second aspect, a mapping relationship between operating status data and load data is established based on the Gaussian process regression algorithm, specifically as follows:

[0034] A machine learning model is trained and established based on the Gaussian process regression algorithm; the machine learning model is used to establish the mapping relationship between operating status data and load data.

[0035] Compared to existing technologies, this invention provides a method and apparatus for establishing a data-driven model. The method includes: establishing a cost minimization objective function and corresponding constraints for a multi-energy park based on relevant parameter data; solving the cost minimization objective function using a deep reinforcement learning algorithm to generate operational status data; establishing a mapping relationship between operational status data and load data using a Gaussian process regression algorithm, and then converting the operational status data into new load data based on the mapping relationship; updating the load database based on the new load data, and establishing a data-driven model for the multi-energy park based on the load database.

[0036] Its beneficial effects are as follows: Based on the relevant parameter data of the multi-energy park, the embodiments of the present invention establish the objective function for minimizing the cost of the multi-energy park and the corresponding constraints. Then, the objective function for minimizing the cost of the multi-energy park is solved by a deep reinforcement learning algorithm to generate operating status data. After establishing the mapping relationship between the operating status data and the load data, the operating status data is converted into new load data according to the mapping relationship. The load database is updated according to the new load data, which can enrich the data volume of the load database. Then, a data-driven model of the multi-energy park is established based on the rich load database, which can effectively avoid the problem of low accuracy of the data-driven model of the multi-energy complementary park generated based on a small amount of load data in the prior art, thereby effectively improving the accuracy of the generated data-driven model of the multi-energy complementary park.

[0037] Furthermore, in this embodiment of the invention, for each type of equipment operating condition change, a deep reinforcement learning algorithm based on multi-agent game theory is used to train agents representing the energy system to obtain coordinated energy management strategies. Then, the operation of each system in the park is simulated in a decentralized manner, optimizing park costs while generating simulated operation data. Then, a machine learning model from park operation data to load data is established based on the Gaussian process regression method, completing the data volume enhancement for multi-energy parks, expanding the sample size, and laying a solid foundation for establishing a data-driven model for multi-energy complementary parks. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a method for establishing a data-driven model according to an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of a data-driven model building device provided in an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Reference Figure 1 This is a flowchart illustrating a method for establishing a data-driven model according to an embodiment of the present invention, including S101-S104:

[0042] S101: Based on the relevant parameter data of the multi-functional park, establish the objective function for minimizing the cost of the multi-functional park and the corresponding constraints.

[0043] Furthermore, the relevant parameter data includes: parameters of each device, typical weather changes, and energy prices. Among them, energy prices include the price of natural gas required for power supply equipment and the time-of-use electricity price of the park; historical electricity and heat load data of similar parks; equipment parameters include the upper and lower limits of the output power of wind power (WT), photovoltaic (PV), power-to-gas (P2G), and gas turbine (GT) equipment, as well as the maximum capacity of the energy storage device (BSS).

[0044] This invention abstracts a multi-energy system as a multi-input, multi-output, multi-port network, referred to as the energy hub. Its input terminals connect to the external power grid / gas network to input electricity / gas, while its output terminals supply energy in the form of electricity, gas, heat, and cooling, meeting the load demand within the park while deriving the optimal energy supply method through global optimization. In this invention, the goal of the multi-energy system optimization process is to find the optimal energy management scheme, obtain operational data under optimized decisions, and complete the data augmentation task. Therefore, the energy management of the multi-energy park needs to be viewed as an optimization-seeking control problem, requiring a detailed description of the system's objective function and constraints.

[0045] In this embodiment, the objective function for minimizing the cost of the multi-functional park is specifically:

[0046] C T =min(C e +C m );

[0047] Among them, C T C represents the total cost of operating a multi-functional park. e Represents operating costs, C m This represents maintenance costs.

[0048] Among them, operating costs The operating costs of the industrial park refer to the expenses incurred in purchasing the necessary energy from outside the park, including the cost of purchasing electricity from the power grid and the cost of purchasing natural gas from the natural gas grid. This can be described as:

[0049]

[0050] in, This represents the amount of electricity input at time t. This represents the amount of natural gas input at time t; This represents the electricity price at time t. This represents the gas price at time t.

[0051] Maintenance costs Maintenance costs, including those for wind power (WT), solar photovoltaic (PV), power-to-gas (P2G), gas turbines (GT), and energy storage systems (BSS), can be described as follows:

[0052]

[0053] in, This represents the output power of the wind power at time t. This represents the output power of photovoltaic power generation at time t. This represents the output power of the gas turbine at time t; This represents the charge level of the energy storage battery at time t. K represents the power consumed by the electro-gas conversion equipment at time t. WT K represents the maintenance cost coefficient for wind power. PV K represents the maintenance cost coefficient for photovoltaic power generation. P2G K represents the maintenance cost coefficient of a gas turbine. GT K represents the maintenance cost coefficient for electro-gas conversion equipment. BSS This represents the maintenance cost coefficient of the energy storage battery.

[0054] The optimization process needs to meet certain constraints, including equipment operation constraints and power balance constraints.

[0055] In this embodiment, the corresponding constraints include:

[0056] Constraints on energy storage batteries, constraints on power-to-gas conversion equipment, constraints on gas turbines, and constraints on the supply and demand balance of natural gas and power systems.

[0057] Among them, the constraints of the energy storage battery: the constraints of the battery storage system at time t can be described as follows:

[0058]

[0059]

[0060]

[0061] in, This represents the charge level of the energy storage battery at time t; This represents the upper limit of the energy storage battery's capacity. This represents the lower limit of the energy storage battery's capacity. This represents the charging power of the BSS at time t-1. η represents the discharge power of the BSS at time t-1; c Indicates charging efficiency, η d Indicates discharge efficiency. This indicates that the energy storage batteries cannot be charged and discharged simultaneously.

[0062] Constraints of electro-gas conversion equipment:

[0063]

[0064]

[0065]

[0066] in, This represents the upper limit of the input power of the electro-gas conversion system. This represents the lower limit of input power for the electro-pneumatic system; It is a binary variable. This indicates that the electro-gas conversion equipment is in operation at time t. This indicates that the electro-gas conversion equipment is in a shutdown state at time t; This indicates that the electro-gas conversion equipment is shut down at time t. This indicates the start-up of the electro-gas conversion equipment at time t.

[0067] Constraints of the gas turbine: The output and ramp-up constraints of the gas turbine are as follows:

[0068]

[0069]

[0070] in, This indicates the lower limit of the gas turbine output. This indicates the upper limit of the gas turbine's output; This represents the downward capability limit of a gas turbine. This represents the limit of the gas turbine's uphill climbing capability.

[0071] Supply and demand balance constraints for natural gas and power systems: For natural gas systems, the sum of natural gas produced by power-to-gas (EPG) units and natural gas purchased from the gas grid must equal the natural gas consumed by the gas turbines; similarly, for power systems, power balance requires that the total electricity produced within the industrial park and purchased from outside the park equal the consumption in each time period. These can be expressed as follows:

[0072]

[0073]

[0074] in, This represents the power consumed by all conventional loads in this energy system at time t. This represents the amount of natural gas stored in the gas tank at time t. This represents the amount of natural gas produced by the electro-gas converter at time t. This represents the amount of natural gas consumed by the gas turbine at time t. This represents the amount of natural gas consumed by other equipment in the park at time t.

[0075] S102: Solve the objective function of minimizing the cost of the multi-functional park using a deep reinforcement learning algorithm to generate operational status data.

[0076] In this embodiment, the step of solving the objective function for minimizing the cost of the multi-functional park using a deep reinforcement learning algorithm to generate operational status data specifically involves:

[0077] Based on the objective function of minimizing the cost of the multi-functional park and the corresponding constraints, an operational status data generation model is generated.

[0078] The running state data generation model is transferred to a Markov game, and the deep reinforcement learning algorithm solves the running state data generation model to calculate the running state data.

[0079] In this embodiment, environmental conditions such as weather conditions, photovoltaic capacity, user load, and energy storage status in the multi-energy park are constantly changing. To achieve the goal of "multi-energy coupling," that is, the dynamic balance between the uncertainties of production and demand in a distributed integrated energy system that includes the coupling and mutual conversion of multiple energy flows (cold, heat, and electricity), the energy management problem is described as a finite-time-domain Markov decision process that does not require prior knowledge of the transition probabilities by defining states, behaviors, rewards, and objective functions. The decision states of each device in the park are recorded at each moment, generating operational status data. The relevant parts are described as follows:

[0080] Data input: A neural network is established for each type of equipment. The environmental information of the multi-energy park (output of photovoltaic equipment, status of energy storage equipment and electricity / heat load demand) is used to read the environmental information of the integrated energy system as the input of the neural network.

[0081] Device state function: Information obtained by each device from the energy hub network composed of multiple inputs and multiple outputs, which can be used to reflect the state of each device. The state set of the entire park at time t is described as follows:

[0082]

[0083] in, This represents the output power of the wind power at time t. This represents the output power of the photovoltaic system at time t. Indicates the charge level of the energy storage battery. This represents the amount of heat energy stored in the thermodynamic system at time t. This represents the selling price of electricity generated in the park at time t. This represents the price of natural gas in the park at time t.

[0084] Equipment action decision a tThis represents the energy management and scheduling decision corresponding to each type of load demand (including electricity, cooling, and heating). After observing the status information within the multi-energy park, each device selects its next action from within the power constraint range. For energy storage batteries, this action is the charging / discharging power at time t. For power-to-gas conversion equipment, its operation is the power consumed at time t; for wind power, gas turbines, and photovoltaic equipment, its operation is the power generated at time t.

[0085]

[0086] Performance feedback: Used to evaluate the performance and effectiveness of device actions and guide each device in the next policy update. During training, all devices share the same reward and train each other to minimize the overall system cost. The reward r is defined. t :

[0087]

[0088]

[0089] in, This represents the total cost of operating the park as defined by the objective function above. This represents the penalty cost incurred when the output power of the energy storage device exceeds the relevant constraints at time t. D represents the penalty coefficient for excessively rapid charging and discharging of the energy storage battery, which is set to a relatively small positive number.

[0090] Operational status data generation: At each moment, each type of device observes and records the status of this system. t Output the action that this type of device should take, a. t During each training session, each device selects a combination of actions based on the overall environmental information of the current park. After this process is completed, the system will transition to the next state. t+1 The system obtains reward and environmental information for the entire park and locally updates the simulated state of each device within the park. This process can be represented as a Markov game. Then, based on the new environmental information, the next round of action selection continues until each agent has learned its own strategy, maximizing the cumulative reward. After the training rounds stabilize, recording the state at each subsequent time step generates optimized operating state curves for each device.

[0091] S=(s n1 s n2 , ..., s nk , ..., s n24 );

[0092]

[0093] Where n represents n iterations, k represents the kth sampling point, and sampling is performed every hour during the day, so k∈[1,24].

[0094] Specifically, taking device i as an example, for each time step t, for a given state s t Using this as the policy input, the π function outputs the probability of each action, and the obtained probabilities are used for random sampling to obtain a. t ~π(·|s t Output the action a that the agent should take. t Get feedback r t and obtain a new state. (s) t a t r t s t+1 The data is stored in a buffer, which holds data under different strategies. Batch sampling (s) is performed from the buffer. i a i r i s i+1 The target value is... State-Action Reward Function: Q(s, a), Given state s, taking action a (a∈A) according to the policy function π yields the expected reward. The environment will provide a corresponding reward r based on the device's action a. Here, γ is the reward discount parameter; a value closer to 1 indicates greater foresight and a focus on the value of subsequent states. Empirically, γ is set to 0.9 here. t This refers to the reward defined earlier. Therefore, the main objective is to construct a Q-value table to store the Q-values ​​of state s and action a, and then select the action a that yields the maximum reward based on the Q-values. The parameters of Q are then updated to make Q(s) equal to... i a i Approximate the value as close to γ ​​as possible. Calculate the loss: L = (Q(s, a) - y) 2 Update Q(s, a) by minimizing the loss. Reset every C steps. Repeat until convergence. After convergence, extract the state s from the Q-value table, where each step contains the current operating state data s of various devices. t :

[0095]

[0096] loop iterations s t →a t →s t+1 →a t+1 →s t+2 …can generate enough runtime status data.

[0097] S103: Based on the Gaussian process regression algorithm, after establishing the mapping relationship between the operating status data and the load data, the operating status data is converted into new load data according to the mapping relationship.

[0098] In this embodiment, establishing the mapping relationship between the operating status data and the load data based on the Gaussian process regression algorithm specifically involves:

[0099] A machine learning model is trained and established based on the Gaussian process regression algorithm; wherein, the machine learning model is used to establish the mapping relationship between the operating status data and the load data.

[0100] The enhanced operational status data includes the output of each device, the energy storage status of the energy storage system, and the prices of electricity and natural gas. This data was derived from training under complex environmental conditions, including changing weather conditions, changing electricity price curves, and demand-load curves from similar industrial parks. Therefore, there is a strong correlation between the enhanced operational status data and the load curves. Since the amount of load data to be expanded in this multi-energy park is relatively small, a Gaussian process regression method suitable for small sample data is adopted.

[0101] The state of the kth sampling point The load at the k-th sampling point is used as input to the Gaussian process regression model. nk For output, the generated load curve is: Load n =(load) n1 load n2 , ..., load n24 ).

[0102] Consider a general noisy regression model load nk =f(s) nk Let f(s) + ε, nk ) comes from a Gaussian process, i.e., f(s) nk )~GP(μ,K), where the noise follows a Gaussian distribution, i.e., ε~N(0,σ) 2 The observed sample dataset is then D = {(s)} n1 load n1 ), (s n2 load n2 ),..,(s n24 load n24 )}, denoted as S n =[s n1 s n2 , ..., s n24 ] T Load n =[load n1 loadn2 , ..., load n24 ] T If these observation points satisfy this Gaussian process, then the joint distribution of these points must follow a Gaussian distribution, that is:

[0103] [f(s n1 ), f(s n2 ), ..., f(s) n24 )] T ~N(μ, K);

[0104] Where, μ=[μ(s) n1 ), μ(s n2 ), ..., μ(s) n24 )] T Let K be the mean vector, and K be a 24*24 dimensional covariance matrix. i，j =k(s ni s nj For a regression problem, given a new running state curve S* = [s 1 *, s 2 *,...,s k *,...,s 24 *] T The goal is to predict loadk. * =f(s) k* )+ε, we can first predict f(S*)=[f(s 1 *), f(s) 2 *), ..., f(s) 24 Then, we can first write the joint distribution of the training data D and the data to be predicted as:

[0105]

[0106] Where, μ(S) n )=μ,μ(S*)=[μ(s 1 *), μ(s) 2 *), ..., μ(s) 24 *)] T ,K(S n S n K(S*, S) = K; K(S*, S) n K(S*, S) is a 24x24 dimensional matrix. n )i,j=k(s i *, s nj K(S*, S*) is a 24*24 dimensional matrix. i，j =k(s i *, s j *). Using the conditional distribution properties of the Gaussian distribution, we can derive:

[0107]

[0108] in,

[0109]

[0110] Adding noise, i.e., letting load* = f(s*) + ε, then

[0111]

[0112] At this point, the load forecast value corresponding to the sample to be predicted can be obtained, i.e., the mean.

[0113] When using Gaussian process regression to predict load from operating state curves, the first step is to select the mean function and the covariance function. The mean function is typically set to 0; the covariance function, also known as the kernel function, is usually a Matern kernel or an SE kernel. The function contains unknown hyperparameters that need to be learned from training samples. Once the hyperparameters are learned, the predicted values ​​can be obtained smoothly. In GPR, hyperparameters are calculated using maximum likelihood estimation, which maximizes the marginal likelihood, i.e., minimizes the negative log marginal likelihood (NLML), as follows:

[0114]

[0115] By iterating multiple times, the hyperparameter θ corresponding to the minimum value of NLML can be found, and the predicted value can be obtained.

[0116] In this scenario, a machine learning model is trained using the Gaussian process regression training method described above, transforming the park's operational status curve into predicted load. When environmental conditions change, the new environmental conditions are first input into the multi-agent reinforcement learning model mentioned above to obtain park operational data. This operational data is then input into the Gaussian process regression model to generate load data for the multi-functional park under the new environmental conditions, thus completing the data augmentation task.

[0117] S104: Update the load database based on the new load data, and establish a data-driven model for the multi-energy park based on the load database.

[0118] In this embodiment, updating the load database based on new load data enriches the data volume of the load database. Then, establishing a data-driven model for multi-energy parks based on the data-rich load database effectively avoids the problem of low accuracy in data-driven models for multi-energy complementary parks generated based on a small amount of load data in the prior art, thereby effectively improving the accuracy of the generated data-driven model for multi-energy complementary parks.

[0119] Furthermore, the new load data added in this invention can serve as the raw data required for the "Multi-Energy Park Load Scenario Boundary Inference Method." After using the additional data provided by this invention, the target method can select a more reasonable uncertainty range based on the degree of impact on the safe operation of the multi-energy complementary park, thereby effectively improving the accuracy of the data-driven model for the multi-energy complementary park. Simultaneously, the new load data added in this invention can also serve as the raw data required for the "Multi-Energy Park Source-Load Scenario Generation Method." After using the additional data provided by this invention, the target method can generate a more representative and extreme set of typical daily scenarios, improving the accuracy of the scenario generation results, and thus providing reference and support for regional integrated energy system planning and scheduling operation schemes.

[0120] Currently, multi-energy complementary industrial parks are mostly demonstration projects, with limited and difficult-to-obtain user-side measurement data samples. During the park planning phase, a pressing issue is how to generate energy efficiency assessment data that meets planning requirements from limited cooling, heating, and power load data. Expanding the small sample planning data of multi-energy parks to establish reasonable data-driven assessment and prediction models is a crucial prerequisite and strong guarantee for the rational planning of multi-energy complementary industrial parks. Using the method of this invention, new load data under corresponding environmental conditions can be generated simply by providing new environmental conditions. This can help establish data-driven models for assessment and prediction of multi-energy parks, providing a data foundation for the planning, design, and assessment of regional multi-energy complementary industrial parks, and promoting the rational operation and development of the urban energy internet.

[0121] To further illustrate the apparatus for building a data-driven model, please refer to [reference needed]. Figure 2 , Figure 2 This is a schematic diagram of a data-driven model building device according to an embodiment of the present invention, including: a first building module 201, a second building module 202, a calculation module 203 and a third building module 204;

[0122] The first establishment module 201 is used to establish the objective function for minimizing the cost of the multi-energy park and the corresponding constraints based on the relevant parameter data of the multi-energy park.

[0123] The second establishment module 202 is used to solve the objective function of minimizing the cost of the multi-functional park according to the deep reinforcement learning algorithm, and generate operation status data;

[0124] The calculation module 203 is used to establish a mapping relationship between the operating status data and the load data according to the Gaussian process regression algorithm, and then convert the operating status data into new load data according to the mapping relationship.

[0125] The third establishment module 204 is used to update the load database according to the new load data, and to establish a data-driven model for the multi-energy park according to the load database.

[0126] In this embodiment, the objective function for minimizing the cost of the multi-functional park is specifically:

[0127] C T =min(C e +C m );

[0128] Among them, C T C represents the total cost of operating a multi-functional park. e Represents operating costs, C m This represents maintenance costs.

[0129] In this embodiment, the corresponding constraints include:

[0130] Constraints on energy storage batteries, constraints on power-to-gas conversion equipment, constraints on gas turbines, and constraints on the supply and demand balance of natural gas and power systems.

[0131] In one specific embodiment, the step of solving the objective function for minimizing the cost of the multi-functional park using a deep reinforcement learning algorithm to generate operational status data specifically involves:

[0132] Based on the objective function of minimizing the cost of the multi-functional park and the corresponding constraints, an operational status data generation model is generated.

[0133] The running state data generation model is transferred to a Markov game, and the deep reinforcement learning algorithm solves the running state data generation model to calculate the running state data.

[0134] In one specific embodiment, establishing the mapping relationship between the operating status data and the load data based on the Gaussian process regression algorithm specifically involves:

[0135] A machine learning model is trained and established based on the Gaussian process regression algorithm; wherein, the machine learning model is used to establish the mapping relationship between the operating status data and the load data.

[0136] In this embodiment of the invention, a first establishment module establishes a cost minimization objective function and corresponding constraints for the multi-energy park based on relevant parameter data; a second establishment module solves the cost minimization objective function using a deep reinforcement learning algorithm to generate operational status data; a calculation module establishes a mapping relationship between operational status data and load data using a Gaussian process regression algorithm, and then converts the operational status data into new load data based on the mapping relationship; a third establishment module updates the load database based on the new load data and establishes a data-driven model for the multi-energy park based on the load database.

[0137] This invention establishes a cost minimization objective function and corresponding constraints for multi-energy parks based on relevant parameter data. Then, it solves the cost minimization objective function using a deep reinforcement learning algorithm to generate operational status data. Next, it establishes a mapping relationship between the operational status data and load data, and converts the operational status data into new load data based on this mapping relationship. The load database is then updated based on the new load data, enriching the data volume of the load database. Finally, a data-driven model for multi-energy parks is established based on this rich load database. This effectively avoids the problem of low accuracy in existing multi-energy complementary park data-driven models generated based on limited load data, thereby significantly improving the accuracy of the generated multi-energy complementary park data-driven model.

[0138] Furthermore, in this embodiment of the invention, for each type of equipment operating condition change, a deep reinforcement learning algorithm based on multi-agent game theory is used to train agents representing the energy system to obtain coordinated energy management strategies. Then, the operation of each system in the park is simulated in a decentralized manner, optimizing park costs while generating simulated operation data. Then, a machine learning model from park operation data to load data is established based on the Gaussian process regression method, completing the data volume enhancement for multi-energy parks, expanding the sample size, and laying a solid foundation for establishing a data-driven model for multi-energy complementary parks.

[0139] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for establishing a data-driven model, characterized in that, include: Based on the relevant parameter data of the multi-functional park, establish the objective function for minimizing the cost of the multi-functional park and the corresponding constraints. The multi-functional park's cost minimization objective function is solved using a deep reinforcement learning algorithm to generate operational status data. After establishing the mapping relationship between the operating status data and the load data according to the Gaussian process regression algorithm, the operating status data is transformed into new load data according to the mapping relationship. The load database is updated based on the new load data, and a data-driven model for the multi-energy park is established based on the load database.

2. The method for establishing a data-driven model according to claim 1, characterized in that, The objective function for minimizing the cost of the multi-functional park is as follows: C T =min(C e +C m ); Among them, C T C represents the total cost of operating a multi-functional park. e Represents operating costs, C m This represents maintenance costs.

3. The method of claim 2, wherein, The corresponding constraints include: Constraints on energy storage batteries, constraints on power-to-gas conversion equipment, constraints on gas turbines, and constraints on the supply and demand balance of natural gas and power systems.

4. The method for establishing a data-driven model according to claim 3, characterized in that, The step of solving the objective function for minimizing the cost of the multi-functional park using a deep reinforcement learning algorithm to generate operational status data specifically involves: Based on the objective function of minimizing the cost of the multi-functional park and the corresponding constraints, an operational status data generation model is generated. The running state data generation model is transferred to a Markov game, and the deep reinforcement learning algorithm solves the running state data generation model to calculate the running state data.

5. The method for establishing a data-driven model according to claim 4, characterized in that, The process of establishing a mapping relationship between the operating status data and the load data based on the Gaussian process regression algorithm is as follows: A machine learning model is trained and established based on the Gaussian process regression algorithm; wherein, the machine learning model is used to establish the mapping relationship between the operating status data and the load data.

6. A data-driven model building apparatus, characterized in that, include: The module consists of a first creation module, a second creation module, a calculation module, and a third creation module. The first establishment module is used to establish the objective function for minimizing the cost of the multi-energy park and the corresponding constraints based on the relevant parameter data of the multi-energy park. The second establishment module is used to solve the objective function of minimizing the cost of the multi-functional park according to the deep reinforcement learning algorithm, and generate operation status data; The calculation module is used to establish a mapping relationship between the operating status data and the load data according to the Gaussian process regression algorithm, and then convert the operating status data into new load data according to the mapping relationship. The third establishment module is used to update the load database based on the new load data and establish a data-driven model for the multi-energy park based on the load database.

7. The data-driven model building apparatus according to claim 6, characterized in that, The objective function for minimizing the cost of the multi-functional park is as follows: C T =min(C e +C m ); Among them, C T C represents the total cost of operating a multi-functional park. e Represents operating costs, C m This represents maintenance costs.

8. The data-driven model building apparatus according to claim 7, characterized in that, The corresponding constraints include: Constraints on energy storage batteries, constraints on power-to-gas conversion equipment, constraints on gas turbines, and constraints on the supply and demand balance of natural gas and power systems.

9. The data-driven model building apparatus according to claim 8, characterized in that, The step of solving the objective function for minimizing the cost of the multi-functional park using a deep reinforcement learning algorithm to generate operational status data specifically involves: Based on the objective function of minimizing the cost of the multi-functional park and the corresponding constraints, an operational status data generation model is generated. The running state data generation model is transferred to a Markov game, and the deep reinforcement learning algorithm solves the running state data generation model to calculate the running state data.

10. The data-driven model building apparatus according to claim 9, characterized in that, The process of establishing a mapping relationship between the operating status data and the load data based on the Gaussian process regression algorithm is as follows: A machine learning model is trained and established based on the Gaussian process regression algorithm; wherein, the machine learning model is used to establish the mapping relationship between the operating status data and the load data.

Images