Start-up optimization method and system for gas storage compressor unit
By combining genetic algorithms and Q-learning reinforcement learning algorithms, the number of compressor units in gas storage facilities is optimized, which solves the high cost problem caused by the same number of compressor units being turned on during different electricity price periods, and realizes the off-peak operation and cost reduction of gas storage compressor units.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PETROCHINA CO LTD
- Filing Date
- 2024-11-05
- Publication Date
- 2026-05-08
AI Technical Summary
In existing technologies, the number of gas storage compressor units that are turned on is the same during different electricity pricing periods, resulting in high operating costs and making it impossible to achieve off-peak operation. Existing genetic algorithm optimization models lack self-learning capabilities and cannot update optimization parameters in a timely manner.
By combining genetic algorithms with Q-learning reinforcement learning, an optimization model is constructed. The mutation rate and crossover rate of the genetic algorithm are adjusted through reinforcement learning to optimize the number of compressor units in the gas storage facility that are turned on at different times.
It enables staggered operation of the gas storage compressor unit, reduces operating costs, optimizes the compressor unit start-up scheme, improves the self-learning ability of the genetic algorithm, and dynamically adjusts the mutation rate and crossover rate to adapt to the optimization needs of different populations.
Smart Images
Figure CN121998145A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and specifically to a method and system for optimizing the start-up of a gas storage compressor unit. Background Technology
[0002] Underground gas storage facilities are artificial gas fields or reservoirs formed by re-injecting natural gas transported through long-distance pipelines into enclosed underground spaces. With the accelerating pace of gas storage facility construction, the optimization of their operation is becoming increasingly important. Large gas storage facilities typically consist of compressor units comprised of numerous compressors used for gas injection and extraction. During operation, electricity costs vary across different time periods (peak, off-peak, and low-cost periods). Currently, the number of compressor units operating is the same regardless of the electricity cost, which is detrimental to cost reduction. Therefore, compressor units need to operate in a staggered manner while ensuring sufficient injection and extraction volumes. This requires a rational planning of the number of compressor units operating during different electricity cost periods, necessitating the research of the lowest-cost compressor unit operation scheme.
[0003] Currently, the optimization of compressor start-up schemes is achieved using genetic algorithms, including single-population and multi-population genetic algorithms. These genetic algorithms can optimize compressor start-up schemes to a certain extent and reduce the operating costs of gas storage compressor units. However, ordinary genetic algorithm optimization models do not have the ability to learn on their own and cannot update the parameters of the optimization algorithm in a timely manner. Therefore, the optimization effect is not optimal, resulting in higher operating costs for compressor units. Summary of the Invention
[0004] To address the high start-up and operation costs of compressor units in existing gas storage facilities, this invention proposes a start-up optimization method and system for compressor units in gas storage facilities. It combines a genetic algorithm and a Q-learning reinforcement learning algorithm to establish an optimization model, optimizing the number of compressor units started at different times during gas storage operation. This solves the problem of high costs caused by the same number of units started during different electricity price periods. This invention allows compressor units to operate off-peak, reducing the operating costs of compressor units in large gas storage facilities.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A startup optimization method for gas storage compressor units specifically includes the following steps:
[0007] S1: Construct an optimization model for the start-up of the gas storage compressor unit based on historical data of the gas storage compressor unit;
[0008] S2: The startup optimization model is trained and optimized based on reinforcement learning and genetic algorithms;
[0009] S3: Input the real-time data of the gas storage compressor unit into the trained start-up optimization model, and output the best start-up plan, i.e. the number of units to be started per hour.
[0010] Preferably, in step S1, the objective function of the startup optimization model is:
[0011]
[0012] In formula (1), The optimization objectives include compressor station electricity costs and start-up losses; n i This represents the number of compressors turned on in the i-th hour; n i-1 This indicates the number of compressors turned on in the (i-1)th hour; P represents the power of a single compressor, in kW; f pi Q represents the electricity price for the i-th hour, in yuan / kW; g This indicates the amount of natural gas emitted when a single compressor starts, in m. 3 ;f g This indicates the price of natural gas, in yuan / m³. 3 .
[0013] Preferably, in step S2, the reinforcement learning algorithm is the Q-learning algorithm.
[0014] Preferably, S2 specifically includes:
[0015] S2-1: Determine the state of reinforcement learning and divide the state into m equal parts;
[0016] S2-2: Determine the reward value for reinforcement learning;
[0017] S2-3: Reinforcement learning optimizes the mutation rate and crossover rate of the genetic algorithm;
[0018] S2-4: Update the Q-table based on the state and reward value of reinforcement learning;
[0019] S2-5: Select an action in the current state using the updated Q-table.
[0020] Preferably, in S2-1, the state of reinforcement learning is:
[0021] S = w1 * f_s + w2 * d_s + w3 * m_s;
[0022]
[0023] In formula (2), S represents the state of reinforcement learning; w1, w2, and w3 represent weight parameters; f_s represents the sum of the fitness of the population to be studied; d_s represents the absolute deviation between the fitness of the population to be studied and the average fitness of the population; m_s represents the maximum fitness of the population to be studied; fitv represents the fitness value of the current generation of the population, and last_fitv represents the fitness value of the previous generation of the population; sum represents the summation function.
[0024] Preferably, in step S2-2, the reward value is determined by the following formula:
[0025]
[0026] In formula (3), r1 represents the reward value when the action is the cross rate value; 2 represents the reward value when the action is the mutation rate.
[0027] Preferably, S2-3 includes:
[0028] S2-3-1: Set the minimum and maximum values of the mutation rate, and divide this interval into n equal parts, where n is the action dimension; each action corresponds to a corresponding small interval.
[0029] S2-3-2: Then, using a loop to traverse each action dimension, the start and end values between each cell are calculated based on the specified minimum and maximum values;
[0030] S2-3-3: Generate floating-point numbers within a selected small interval using a random method, and use them as the mutation rate for the corresponding action dimension.
[0031] Preferably, in steps S2-4, the Q table update method is as follows:
[0032] Q(s,a)=Q(s,a)+α[R+γmax a′ Q(s′,a′)-Q(s,a)] (4)
[0033] In formula (4), Q(s,a) represents the Q value of the current action a in the current state s; α represents the learning rate; R represents the reward value; γ represents the decay rate; and Q(s′,a′) represents the Q value of the next action a′ in the next state s′.
[0034] Preferably, in step S3, the constraints in the real-time data of the gas storage compressor unit are as follows:
[0035]
[0036] In formula (5), n i This represents the number of compressors started in the i-th hour; q represents the injection volume per hour of a single compressor, in ten thousand cubic meters per hour; Q z This indicates the total daily gas injection volume required by the compressor, in ten thousand cubic meters per day.
[0037] The present invention also provides a start-up optimization system for a gas storage compressor unit, comprising:
[0038] The data acquisition unit is used to acquire historical data and real-time data of the gas storage compressor unit from the server.
[0039] The model building unit is used to build an start-up optimization model based on historical data of the gas storage compressor unit.
[0040] The model training unit is used to train and optimize the startup optimization model;
[0041] The output unit is used to output the start-up plan, i.e. the number of start-ups per hour, based on the real-time data of the gas storage compressor unit and the trained start-up optimization model.
[0042] In summary, by adopting the above technical solution, the present invention has at least the following beneficial effects compared with the prior art:
[0043] This invention establishes an optimization model by combining genetic algorithms and Q-learning reinforcement learning algorithms, thereby optimizing the number of compressor units that are started at different times during the operation of a gas storage facility. This solves the problem of high costs caused by the same number of units being started during different electricity price periods. This invention allows compressor units to operate in shifts, reducing the operating costs of compressor units in large gas storage facilities. Attached Figure Description
[0044] Figure 1 This is a schematic diagram of a start-up optimization method for a gas storage compressor unit according to an exemplary embodiment of the present invention.
[0045] Figure 2 This is a schematic diagram comparing the number of devices started per hour according to an exemplary embodiment of the present invention.
[0046] Figure 3 This is a schematic diagram illustrating a daily cost comparison according to an exemplary embodiment of the present invention.
[0047] Figure 4 This is a schematic diagram of a start-up optimization system for a gas storage compressor unit according to an exemplary embodiment of the present invention. Detailed Implementation
[0048] The present invention will be further described in detail below with reference to embodiments and specific implementation methods. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.
[0049] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0050] like Figure 1 As shown, the present invention provides a start-up optimization method for a gas storage compressor unit, specifically including the following steps:
[0051] S1: Construct an optimization model for the start-up of the gas storage compressor unit based on historical data of the gas storage compressor unit.
[0052] In this embodiment, the start-up scheme for the large gas storage compressor unit needs to minimize operating costs. Operating costs include the compressor station electricity cost (i.e., the cost of running the compressor) and start-up loss costs (i.e., the cost of natural gas emissions during compressor startup). Therefore, the start-up optimization model should aim to minimize the sum of the compressor station electricity cost and the start-up loss costs. The objective function is:
[0053]
[0054] In formula (1), The optimization objectives include compressor station electricity costs and start-up losses; n i This represents the number of compressors turned on in the i-th hour; n i-1 This indicates the number of compressors turned on in the (i-1)th hour; P represents the power of a single compressor, in kW; f pi Q represents the electricity price for the i-th hour, in yuan / kW; g This indicates the amount of natural gas emitted when a single compressor starts, in m. 3 ;f g This indicates the price of natural gas, in yuan / m³. 3 .
[0055] S2: The startup optimization model is trained and optimized based on reinforcement learning and genetic algorithms.
[0056] In this embodiment, the Q-learning reinforcement learning algorithm is selected. The core of this algorithm is to use the Q-table to select the action with the largest Q-value in the current state and then proceed to the next state. Therefore, there are three key points to choosing Q-learning for reinforcement learning: 1. The next state after selecting an action in the current state; 2. The reward value for selecting an action in a given state; 3. How to select an action in the current state. Since it is known to be based on the Q-table, the two main tasks are updating the Q-table and using the Q-table to select an action in the current state.
[0057] After selecting a reinforcement learning algorithm, it's necessary to determine the state dimension and action dimension. The state dimension is the number of generations in the population, and the action dimension can be set. (The state dimension s...) m and action dimension a n Create an initial Q-table with rows representing the state dimension and columns representing the action dimension. In this problem, two Q-tables need to be created because there are two actions to choose from: mutation rate and crossover rate.
[0058] Table 1. Initial Q Table
[0059] <![CDATA[a1]]> <![CDATA[a2]]> <![CDATA[a3]]> ... <![CDATA[a n ]]> <![CDATA[s1]]> <![CDATA[Q(s1,a1)]]> <![CDATA[Q(s1,a2)]]> <![CDATA[Q(s1,a3)]]> <![CDATA[Q(s1,a n )]]> <![CDATA[s2]]> <![CDATA[Q(s2,a1)]]> <![CDATA[Q(s2,a2)]]> <![CDATA[Q(s2,a3)]]> <![CDATA[Q(s2,a n )]]> ... <![CDATA[s m ]]> <![CDATA[Q(s m ,a1)]]> <![CDATA[Q(s m a2)]]> <![CDATA[Q(s m a3)]]> <![CDATA[Q(s m ,a n )]]>
[0060] S2-1: Determine the state of reinforcement learning and divide the state into m equal parts, where m is the state dimension and the corresponding state corresponds to the next state.
[0061] In this embodiment, the state value of reinforcement learning cannot be determined solely based on the fitness value of the current generation. This is because it needs to reflect the comparison between the fitness value of the next generation and the fitness value of the previous generation. Therefore, the state value is related not only to the fitness of the current generation but also to the fitness of the previous generation. Moreover, when considering fitness, the sum, maximum, and average fitness of the first generation are all very important. The sum of fitness covers the fitness of all individuals in the population, making the results more convincing. The maximum fitness reflects the highest fitness in the population, which is the fitness value that is of most concern in genetic algorithms. The core of genetic algorithms is optimization, which is to obtain the phenotype of the individual with the highest fitness.
[0062] The state of reinforcement learning is related to the current generation fitness and the previous generation fitness of the genetic algorithm population. The core is the objective function of the genetic algorithm: the operating cost of the gas storage compressor unit, which is determined by the following formula:
[0063] S = w1 * f_s + w2 * d_s + w3 * m_s;
[0064]
[0065] In formula (2), S represents the state of reinforcement learning; w1, w2, and w3 represent weight parameters; f_s represents the sum of the fitness of the population to be studied; d_s represents the absolute deviation between the fitness of the population to be studied and the average fitness of the population; m_s represents the maximum fitness of the population to be studied; fitv represents the fitness value of the current generation of the population, last_fitv represents the fitness value of the previous generation of the population; and sum represents the summation of the fitness values of all individuals in the population.
[0066] S2-2: Determine the reward value for reinforcement learning.
[0067] The reward in reinforcement learning is a function related to the current fitness and the previous generation fitness of the population in the genetic algorithm. The current fitness value is determined by the previous generation fitness value, and the core is the objective function of the genetic algorithm (the operating cost of the gas storage compressor unit).
[0068] In a given state, choosing an action to proceed to the next state results in two reward values because there are two Q-tables. The reward value is determined by the following formula.
[0069]
[0070] In formula (3), r1 represents the reward value when the action is the cross rate value; 2 represents the reward value when the action is the mutation rate.
[0071] S2-3: Reinforcement learning optimizes the mutation rate and crossover rate of the genetic algorithm.
[0072] In a typical genetic algorithm, the mutation rate and crossover rate remain constant across generations. However, for different generations, different mutation and crossover rates obviously have a higher probability of evolving individuals with higher fitness. Since the changes in mutation and crossover rates are not random, reinforcement learning can be used to select the mutation and crossover rates that result in higher fitness in each generation for further evolution.
[0073] S2-3-1: The actions in reinforcement learning are the mutation rate and crossover rate of the population. In order to establish the relationship between the actions in reinforcement learning and the mutation rate and crossover rate of the genetic algorithm, we can set the minimum and maximum values of the mutation rate and divide this interval into n equal parts, where n is the action dimension; each action corresponds to a corresponding small interval.
[0074] S2-3-2: Then, using a loop to traverse each action dimension, calculate the start and end values between each cell based on the specified minimum and maximum values.
[0075] S2-3-3: Next, a floating-point number within a selected small interval is generated using a random method, which serves as the mutation rate for the corresponding action dimension.
[0076] Similarly, the cross rate is handled in a similar way, except that the minimum and maximum values are set differently.
[0077] In this way, the action space of reinforcement learning can be linked to the mutation rate and crossover rate of the genetic algorithm. This allows the genetic algorithm to control the search range and diversity of the solution space by adjusting the mutation rate and crossover rate, thereby further optimizing the policy search process of reinforcement learning.
[0078] S2-4: Update the initial Q table.
[0079] In this embodiment, the Q values are initialized to all 0 when the Q table is created, and the Q values need to be updated and calculated using a formula:
[0080] Q(s,a)=Q(s,a)+α[R+γmax a′ Q(s′,a′)-Q(s,a)] (4)
[0081] In formula (4), Q(s,a) represents the Q value of the current action a in the current state s; α represents the learning rate; R represents the reward value, including r1 and r2; γ represents the decay rate; and Q(s′,a′) represents the Q value of the next action a′ in the next state s′.
[0082] The clever part of updating the initial Q-table is that it not only considers the estimated value of Q(s,a) (the value before the update), but also the maximum Q-value of all selectable actions after performing action a in state s and reaching the next state s′. In other words, it considers both the current state and the next step, but the Q-value of the next state's action is not fully adopted; instead, it is accompanied by a decay coefficient. The reward value here is generally 0 until the final goal is reached.
[0083] S2-5: Select an action in the current state using the updated Q-table.
[0084] Epsilon greedy is a strategy used in decision-making. epsilon is the greedy parameter of the Q-learning algorithm. Using a random function, it selects the action according to the optimal value of the Q-table in the case of 1-epsilon, and randomly selects the action in the case of epsilon.
[0085] In this embodiment, "state" and "action" are concepts in reinforcement learning, while "mutation rate" and "crossover rate" are concepts in genetic algorithms. The innovative approach of this genetic algorithm is to use reinforcement learning to dynamically adjust the mutation rate and crossover rate. These two values are changed because adaptive updates to these values lead to better performance from the genetic algorithm. This innovative genetic algorithm is used to solve the S1 objective function.
[0086] Genetic algorithms are used to solve a problem given an objective function (S1) containing decision variables (number of machines started per hour) and constraints.
[0087] S3: Input the real-time data of the gas storage compressor unit into the trained start-up optimization model, and output the best start-up plan, i.e. the number of units to be started per hour.
[0088] In this embodiment, the real-time data of the gas storage compressor unit includes compressor power, electricity price, natural gas emission from a single compressor startup, natural gas price, hourly injection volume of a single compressor, and daily required gas injection volume.
[0089] For example, the electricity price for a gas storage compressor unit 24 hours a day. pi for:
[0090] Peak hours: 8:00-12:00 and 19:00-23:00 (8 hours total), electricity price: 0.7479 yuan / kWh. Off-peak hours: 7:00-8:00 and 12:00-19:00 (8 hours total), electricity price: 0.4986 yuan / kWh. Low-peak hours: 23:00 to 7:00 the next day (8 hours total), electricity price: 0.2493 yuan / kWh.
[0091] —The power P of a single compressor is 3600kw, and the amount of natural gas emitted by a single compressor when it starts is Q. g 1500m 3 Natural gas price f g It is 2.83 yuan / m 3 The injection rate q of a single compressor is 79,000 cubic meters per hour, and the daily required gas injection volume Q is... z The capacity is 8.8 million cubic meters per day, and the number of compressors is n. i It is 6.
[0092] In this embodiment, the decision variable of the enhanced genetic algorithm is the number of compressors started each hour within a 24-hour period. The decision variable is discrete and has a dimension of 24. The optimization objective is the operating cost of the large gas storage compressor unit, with an objective dimension of 1, and the optimization direction is to minimize the objective cost.
[0093] Genetic algorithms are used to solve a problem given an objective function (S1) containing decision variables (number of machines started per hour) and constraints.
[0094] Therefore, the reinforced genetic algorithm in this project has two constraints: first, the number of compressors cannot exceed the maximum number of compressors; second, the total daily gas injection and extraction volume must meet the daily total gas injection and extraction volume requirement. After clarifying these parameters and variables, the reinforced genetic algorithm can be used to solve for the optimal start-up scheme of the gas storage compressor unit.
[0095] The parameter constraints in the objective function are as follows:
[0096] Compressor quantity constraint: the number of compressors in operation must be within the total number of compressors in the compressor station.
[0097] n i ∈[1,2,...,6]
[0098] Gas injection volume constraint: the total daily gas injection volume meets the total daily gas demand requirement;
[0099]
[0100] In formula (5), q represents the injection volume of a single compressor per hour, in ten thousand cubic meters per hour; Q z This indicates the total daily gas injection volume required by the compressor, in ten thousand cubic meters per hour.
[0101] This invention combines reinforcement learning and genetic learning algorithms. Ordinary genetic algorithms cannot achieve self-updating of the mutation rate and crossover rate. For different populations, different mutation rates and crossover rates obviously have a higher probability of evolving individuals with higher fitness. Therefore, to improve the performance of genetic algorithms, a reinforcement learning-based genetic algorithm is used to dynamically adjust the mutation rate and crossover rate. Using reinforcement learning, we can select mutation rates and crossover rates that yield higher fitness in each generation of the population, achieving self-updating of the mutation rate and crossover rate, and using these for the evolution of the next generation of the population to achieve population self-adaptation. The reinforcement learning-based genetic algorithm can better optimize the start-up scheme of large gas storage compressor units, thereby achieving lower operating costs and reducing compressor operating expenses. This invention can not only be applied to the optimization of start-up schemes for large gas storage compressors, but also to other fields requiring parameter optimization, and has significant promotional value.
[0102] To verify the technical solution of this invention, the optimized boot scheme and the initial boot scheme are compared, such as... Figure 2As shown. Under the initial startup scheme, the number of compressors turned on is fixed at 5 units within 24 hours; under the single-population genetic algorithm, 6 units are turned on from 0:00 to 8:00, 0 units are turned on from 8:00 to 12:00, 4 units are turned on from 12:00 to 13:00, 6 units are turned on from 13:00 to 14:00, and 6 units are turned on from 14:00 to 24:00; under the multi-population genetic algorithm, 6 units are turned on from 0:00 to 8:00, 2 units are turned on from 8:00 to 11:00, 4 units are turned on from 11:00 to 12:00, 5 units are turned on from 12:00 to 18:00, and 4 units are turned on from 18:00 to 24:00; under the technical solution of this invention, 6 units are turned on from 0:00 to 6:00, 4 units are turned on from 6:00 to 7:00, 4 units are turned on from 7:00 to 23:00, and 6 units are turned on from 23:00 to 24:00.
[0103] like Figure 3 As shown, comparing the optimized startup scheme with the initial startup scheme, the daily operating cost of the compressor has been significantly reduced.
[0104] like Figure 4 As shown, based on the above-mentioned startup optimization method for a gas storage compressor unit, the present invention also provides a startup optimization system for a gas storage compressor unit, including a data acquisition unit, a model building unit, a model training unit, and an output unit; the output end of the data acquisition unit is connected to the input end of the model building unit, the output end of the model building unit is connected to the input end of the model training unit, and the output end of the model training unit is connected to the input end of the output unit.
[0105] The data acquisition unit is used to acquire historical data and real-time data of the gas storage compressor unit from the server.
[0106] The model building unit is used to build an start-up optimization model based on historical data of the gas storage compressor unit.
[0107] The model training unit is used to train and optimize the startup optimization model;
[0108] The output unit is used to output the start-up plan, i.e. the number of start-ups per hour, based on the real-time data of the gas storage compressor unit and the trained start-up optimization model.
[0109] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.
Claims
1. A start-up optimization method for a gas storage compressor unit, characterized in that, Specifically, the following steps are included: S1: Construct an optimization model for the start-up of the gas storage compressor unit based on historical data of the gas storage compressor unit; S2: The startup optimization model is trained and optimized based on reinforcement learning and genetic algorithms; S3: Input the real-time data of the gas storage compressor unit into the trained start-up optimization model, and output the best start-up plan, i.e. the number of units to be started per hour.
2. The start-up optimization method for a gas storage compressor unit as described in claim 1, characterized in that, In S1, the objective function of the startup optimization model is: In formula (1), The optimization targets include compressor station electricity costs and start-up losses. n i This represents the number of compressors that are turned on in the i-th hour; n i-1 This indicates the number of compressors that are turned on in the (i-1)th hour; P represents the power of a single compressor, in kW; f pi Q represents the electricity price for the i-th hour, in yuan / kW; g This indicates the amount of natural gas emitted when a single compressor starts, in m. 3 ; f g This indicates the price of natural gas, in yuan / m³. 3 .
3. The start-up optimization method for a gas storage compressor unit as described in claim 1, characterized in that, In S2, the reinforcement learning algorithm is the Q-learning algorithm.
4. The start-up optimization method for a gas storage compressor unit as described in claim 1, characterized in that, S2 specifically includes: S2-1: Determine the state of reinforcement learning and divide the state into m equal parts; S2-2: Determine the reward value for reinforcement learning; S2-3: Reinforcement learning optimizes the mutation rate and crossover rate of the genetic algorithm; S2-4: Update the Q-table based on the state and reward value of reinforcement learning; S2-5: Select an action in the current state using the updated Q-table.
5. The start-up optimization method for a gas storage compressor unit as described in claim 4, characterized in that, In S2-1, the state of reinforcement learning is: S = w1 * f_s + w2 * d_s + w3 * m_s; In formula (2), S represents the state of reinforcement learning; w1, w2, and w3 represent weight parameters; f_s represents the sum of the fitness of the population to be studied; d_s represents the absolute deviation between the fitness of the population to be studied and the average fitness of the population; m_s represents the maximum fitness of the population to be studied; fitv represents the fitness value of the current generation of the population, and last_fitv represents the fitness value of the previous generation of the population; sum represents the summation function.
6. The start-up optimization method for a gas storage compressor unit as described in claim 4, characterized in that, In S2-2, the reward value is determined by the following formula: In formula (3), r1 represents the reward value when the action is the cross rate value; 2 represents the reward value when the action is the mutation rate.
7. The start-up optimization method for a gas storage compressor unit as described in claim 4, characterized in that, S2-3 includes: S2-3-1: Set the minimum and maximum values of the mutation rate, and divide this interval into n equal parts, where n is the action dimension; each action corresponds to a corresponding small interval; S2-3-2: Then, using a loop to traverse each action dimension, the start and end values between each cell are calculated based on the specified minimum and maximum values; S2-3-3: Generate floating-point numbers within a selected small interval using a random method, and use them as the mutation rate for the corresponding action dimension.
8. The start-up optimization method for a gas storage compressor unit as described in claim 4, characterized in that, In S2-4, the Q table update method is as follows: Q(s,a)=Q(s,a)+α[R+γmax a′ Q(s ′ ,a ′ )-Q(s,a)] (4) In formula (4), Q(s,a) represents the Q value of the current action a in the current state s; α represents the learning rate; R represents the reward value; γ represents the decay rate; Q(s ′ ,a ′ ) represents the next state s ′ Next action a ′ The Q value.
9. The start-up optimization method for a gas storage compressor unit as described in claim 1, characterized in that, In S3, the constraints in the real-time data of the gas storage compressor unit are as follows: In formula (5), n i This represents the number of compressors started in the i-th hour; q represents the injection volume per hour of a single compressor, in ten thousand cubic meters per hour; Q z This indicates the total daily gas injection volume required by the compressor, in ten thousand cubic meters per day.
10. A start-up optimization system for a gas storage compressor unit based on the method of any one of claims 1-9, characterized in that, include: The data acquisition unit is used to acquire historical data and real-time data of the gas storage compressor unit from the server. The model building unit is used to build an start-up optimization model based on historical data of the gas storage compressor unit. The model training unit is used to train and optimize the startup optimization model; The output unit is used to output the start-up plan, i.e. the number of start-ups per hour, based on the real-time data of the gas storage compressor unit and the trained start-up optimization model.