A neural network-based parameter optimization method for soft start cabinet of electric submersible pump

Abstract

The application belongs to the technical field of artificial intelligence and data processing, and particularly relates to a neural network-based parameter optimization method for a submersible electric pump soft start cabinet. The method comprises the following steps: obtaining downhole environment characteristics, historical state characteristics and power grid environment characteristics of the submersible electric pump, performing normalization processing and constructing a state vector; constructing a double neural network model based on an actor-critic architecture, using an actor network to establish a mapping from the state to the soft start control parameter, and using a critic network to evaluate the action value; constructing a comprehensive reward function comprising a start result indicator factor, a current peak value constraint and a time constraint; using a deep deterministic policy gradient algorithm, combining an experience replay mechanism and the comprehensive reward function to iteratively train the double neural network model, and generating optimal soft start control parameters. The application can adaptively adjust the start parameters according to real-time working conditions, ensure the successful start of the submersible electric pump, and significantly reduce the impact of the start current on a weak power grid.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data processing technology, specifically relating to a method for optimizing the parameters of a submersible electric pump soft starter based on neural networks. Background Technology

[0002] As a core lifting device in the oil extraction field, the ESP (Electric Submersible Pump) is widely used in various oilfield operations due to its high lifting capacity, playing an irreplaceable role, especially in special operating areas such as the heavy oil block of Shengli Oilfield and offshore platforms. However, the power supply systems in these special operating areas have significant weak grid characteristics, namely, they often use independent microgrids or long-distance transmission lines for power supply, resulting in small short-circuit capacity, high system impedance, relatively weak grid stability, and poor tolerance to impact loads.

[0003] In the startup control of submersible electric pumps, the industry currently widely uses thyristor soft starter cabinets as the core control equipment. The setting methods for their output parameters, such as initial voltage, ramp-up time, and current limiting setpoints, are mainly divided into two categories: manual experience-based setting and fixed value mode. However, in actual production conditions, both of these setting methods face significant challenges:

[0004] On the one hand, downhole operating conditions have extremely strong nonlinear characteristics. The viscosity of crude oil fluctuates drastically with factors such as the temperature of the fluid at the bottom of the well and the duration of downtime, which in turn leads to significant changes in the starting resistance torque of the submersible pump. If fixed low parameters are used for starting, the motor is prone to stalling due to insufficient output torque, affecting the continuity of production operations. If the starting parameters are blindly increased to overcome the high resistance torque, a huge surge current will be generated. In a weak power grid environment, this current will cause a momentary drop in bus voltage, thereby triggering the undervoltage protection mechanism. In severe cases, it may even cause the entire platform equipment to shut down, resulting in significant production losses.

[0005] On the other hand, as the oilfield exploitation cycle progresses, formation parameters will continue to evolve dynamically, and the performance of submersible electric pumps may also change due to factors such as aging and wear, making it difficult for the initially set fixed parameters to adapt to the operating conditions throughout the entire life cycle. Manually adjusting parameters is not only inefficient but also has a significant lag, failing to respond promptly to real-time changes in operating conditions. This increases maintenance costs and makes it difficult to guarantee the safety and stability of the startup process. Summary of the Invention

[0006] The purpose of this invention is to propose a parameter optimization method for the soft starter cabinet of a submersible electric pump based on neural networks, in order to solve the technical problem that the existing technology cannot adapt to nonlinear operating conditions and is prone to power grid impact when using fixed parameters.

[0007] The technical solution of the submersible electric pump soft starter parameter optimization method based on neural network provided by this invention is as follows:

[0008] A method for optimizing the parameters of a submersible electric pump soft starter based on neural networks includes the following steps:

[0009] Acquire downhole environmental characteristic data, historical state characteristic data and power grid environmental characteristic data of submersible electric pumps, and construct a state vector representing the current operating condition after normalizing the acquired data;

[0010] A dual neural network model based on an actor-critic architecture is constructed; wherein, the actor network is used to output the corresponding action policy according to the state vector, and the action policy includes soft-start control parameters; the critic network is used to evaluate the expected value of the action policy;

[0011] A comprehensive reward function is constructed, which includes a startup result indicator factor, a peak current ratio, and a time constraint term. The comprehensive reward function is configured as follows: when the submersible electric pump starts successfully, a positive reward is calculated based on the peak current ratio using a power function; when the startup fails, a negative penalty is calculated based on the stall duration using an exponential function; and the feedback reward value for a single startup process is output in combination with the time constraint term.

[0012] The dual neural network model is iteratively trained using a deep deterministic policy gradient algorithm, through an experience replay mechanism and the feedback reward value; the state vector is input into the executor network after training convergence to generate the optimal soft-start control parameters for the current operating condition.

[0013] This invention establishes a deep mapping relationship between complex nonlinear downhole conditions and optimal startup parameters by constructing a dual neural network model based on an executor-commentator architecture. It utilizes the powerful fitting ability of neural networks to replace traditional manual experience rules, achieving adaptive generation of startup parameters. At the same time, guided by a comprehensive reward function, the system can automatically find a balance between ensuring successful startup and reducing current surges, effectively solving the problems of stalling or grid voltage drop caused by parameter mismatch in traditional methods.

[0014] Furthermore, the downhole environmental characteristic data includes bottom hole fluid temperature data and bottom hole fluid pressure data; the historical state characteristic data includes current downtime data; and the power grid environmental characteristic data includes real-time power grid bus voltage data and power grid frequency data; the state vector satisfies the expression:

[0015]

[0016] In the formula, For normalization function, The temperature of the fluid at the bottom of the well. The bottom hole fluid pressure, This is the current downtime. This represents the real-time voltage of the power grid bus. This refers to the power grid frequency.

[0017] The current downtime can indirectly reflect the degree of crude oil viscosity recovery, and the power grid parameters can reflect the current stability margin. Combining these characteristic data can accurately characterize the current start-up load characteristics and environmental constraints, providing a comprehensive and accurate input basis for the neural network's decision-making.

[0018] Furthermore, the soft-start control parameters include initial voltage percentage, ramp-up time, and current limiting setting; the output layer of the actuator network adopts a hyperbolic tangent activation function, and the output result is mapped to the device safety threshold range through linear transformation;

[0019] The safety threshold range of the device is: initial voltage percentage of 30% to 80%, ramp-up time of 5s to 60s, and current limiting setting of 2.0 to 5.0 times the rated current.

[0020] Furthermore, the comprehensive reward function satisfies the expression:

[0021]

[0022] In the formula, These are the weighting coefficients. As an indicator of the start-up results, Rated current, This represents the measured peak current. The current sensitivity index, For the duration of the stall, To create a penalty index for traffic jams, This is the actual startup time. For ideal startup time, This is for feedback and reward values.

[0023] When startup is successful, the power function characteristic is used to guide the peak current to approach the rated value. When startup fails, the exponential function is used to severely punish the stall time. This design forces the neural network to avoid high-risk areas during the exploration process, thereby training a control strategy that is both safe and efficient.

[0024] Furthermore, the initiation result indicator factor The determination method is as follows: if the motor current is detected to drop back to within a preset multiple of the rated current and the speed meets the standard, then The value is 1; if stall protection, overcurrent protection, or timeout prevents the start-up from being completed, then... The value is 0.

[0025] Furthermore, the iterative training process of the dual neural network model includes: updating the critic network parameters by minimizing the loss function, and minimizing the loss function. Satisfying the expression:

[0026]

[0027] In the formula, The size of the sampled batch of data. For the first in the batch data Index of each sample, The current state. For the action at the current moment, For the critic's network parameters, For the critic network in parameters Next state and actions The evaluation value, This is the target value.

[0028] Furthermore, the target value Satisfying the expression:

[0029]

[0030] In the formula, For the first The feedback reward value for each sample, As a discount factor, For the state at the next moment, and These are the critic target network and the actor target network, respectively. and These are the network parameters for the critic and the network parameters for the executor, respectively. For the executor target network in parameters The recommended best action to take in the next state.

[0031] Furthermore, the gradient algorithm utilizing a deep deterministic policy includes the following steps:

[0032] Establish an experience replay pool. At each time step, store a quadruple sample containing the current state, the action performed, the feedback value, and the state at the next time step into the experience replay pool.

[0033] Batch data is randomly sampled from the experience replay pool to update the network parameters of the executor network and the critic network, and then the corresponding target network parameters are updated using a soft update method.

[0034] Furthermore, the state vector is input into the converged executor network to generate optimal soft-start control parameters for the current operating condition, including:

[0035] When the submersible electric pump is about to start, real-time data is collected to construct the current state vector, which is then input into the trained executor network to obtain the optimal action.

[0036] The soft starter executes the optimal action to control the thyristor conduction angle, and calculates the feedback value after startup. The startup data is then stored in the experience playback pool for subsequent incremental training.

[0037] Furthermore, both the executor network and the critic network contain two fully connected hidden layers; wherein the first fully connected hidden layer has 256 nodes and the second fully connected hidden layer has 128 nodes, and both hidden layers use the ReLU activation function.

[0038] The beneficial effects of this invention are as follows: By constructing a dual neural network model based on a deep deterministic strategy gradient algorithm, this invention establishes a deep mapping relationship between complex nonlinear downhole conditions and optimal startup parameters. The powerful fitting ability of the neural network replaces traditional manual experience rules, achieving adaptive generation of startup parameters. Simultaneously, this invention designs a comprehensive reward function. Upon successful startup, the power function characteristic guides the peak current to approach the rated value; upon startup failure, an exponential function severely penalizes the stall time. The system can automatically find a balance between ensuring successful startup and reducing current surges, effectively solving the stall or grid voltage drop problems caused by parameter mismatch in traditional methods. Furthermore, this invention adopts a closed-loop control mode of offline pre-training and online correction. The system can store the data from each actual startup as experience in a playback pool for incremental learning, enabling the system to have full lifecycle self-evolution capabilities. It can automatically adapt to the impact of equipment aging or changes in the formation environment, significantly reducing operation and maintenance costs. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the steps of a method for optimizing the parameters of a submersible electric pump soft starter based on a neural network, according to the present invention.

[0040] Figure 2 This is a schematic diagram comparing the starting current waveforms of the submersible electric pump in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0042] This invention discloses a neural network-based method for optimizing the soft starter parameters of a submersible electric pump. This method, based on a deep deterministic strategy gradient algorithm, utilizes a dual neural network model with an executor-commentator architecture to achieve an end-to-end mapping from complex downhole conditions to optimal soft starter control parameters. For example... Figure 1 As shown, steps S1-S4 are included:

[0043] S1. Obtain downhole environmental characteristic data, historical state characteristic data, and power grid environmental characteristic data of the submersible electric pump. After normalizing the acquired data, construct a state vector representing the current operating condition.

[0044] Specifically, in order for the neural network to perceive the current operating environment, it is necessary to collect multi-dimensional physical quantities. Downhole environmental characteristic data includes the bottomhole fluid temperature, which is collected in real time by downhole sensor units installed at the bottom of the submersible motor. and bottom hole fluid pressure These two parameters directly determine the viscosity characteristics of crude oil, thus affecting the starting resistance torque. Historical status characteristic data includes the current downtime read from the soft starter PLC controller. The longer the downtime, the stronger the thixotropy of the crude oil in the wellbore, and the greater the initial torque required for startup. Power grid environmental characteristic data includes real-time voltage of the power grid bus collected by incoming line power meters. and grid frequency This is used to assess the voltage support capability under weak power grid conditions. To eliminate the influence of different physical dimensions on the neural network gradient, the collected data needs to be processed by min-max normalization. At this point, the state vector satisfies the expression:

[0045]

[0046] In the formula, The normalization function is calculated using the following formula: ,in and Based on the device nameplate and historical extreme values.

[0047] For example, assume the collected real-time data and parameter reference ranges are as follows:

[0048] Bottom hole fluid temperature The reference range is Bottom hole fluid pressure The reference range is Current downtime Hours, reference range is Hours; Grid voltage The reference range is ; power grid frequency The reference range is .

[0049] The calculation results for each component of the state vector are as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] .

[0055] The final constructed current moment The state vector is: .

[0056] This step, through the collection and normalization of multi-source heterogeneous data, can transform complex downhole physical conditions and power grid fluctuations into standardized input vectors that can be processed by neural networks, providing a precise and unified data foundation for subsequent intelligent decision-making.

[0057] S2. Construct a dual neural network model based on an executor-critic architecture; wherein, the executor network is used to output the corresponding action policy according to the state vector, and the action policy includes soft-start control parameters; the critic network is used to evaluate the expected value of the action policy.

[0058] Specifically, this step constructs a dual neural network structure based on the deep deterministic policy gradient algorithm. Both the executor network and the critic network contain two fully connected hidden layers; the first fully connected hidden layer has 256 nodes, the second fully connected hidden layer has 128 nodes, and both fully connected hidden layers use the ReLU activation function.

[0059] The executor network is responsible for fitting the policy function. Its input layer receives the state vector. Features are extracted through two fully connected hidden layers. The output layer of the actuator network uses a hyperbolic tangent activation function to restrict the output value to the range of [-1, 1], and further maps the output value to the device safety threshold range through a linear transformation. The output layer of the actuator network outputs three soft-start control parameters, namely the initial voltage percentage. Slope ascent time and current limiting settings The equipment safety threshold range is as follows: initial voltage percentage from 30% to 80%, ramp-up time from 5s to 60s, and current limiting setting from 2.0 to 5.0 times the rated current.

[0060] The formula for linear transformation is: .

[0061] For example, suppose the actor network is concerned with the state in step S1 The output original action vector is The corresponding actual soft-start control parameters are calculated as follows:

[0062] Initial voltage percentage ;

[0063] Slope ascent time ;

[0064] Current limiting setting Ultimately, the soft starter will start with 42.5% of the initial voltage percentage, a ramp-up time of 38 seconds, and 4.7 times the current limit setting.

[0065] The critic network is responsible for fitting the value function. The input layer receives the state vector. and the action vector output by the executor network The output is a scalar. The value is used to characterize the expected value of the action in the current state, that is, the state at the current moment. If action is taken How much benefit can it bring in the future? The output scalar. The value is used to guide the updates of the executor network, causing the executor network to follow the scalar... Adjust the neural network parameters in the direction that the value increases.

[0066] Thus, through the executor-commentator architecture and linear mapping of physical space, the system can directly generate precise control parameters in the continuous action space, avoiding the precision limitations of traditional discrete-level control and realizing fine-grained control of the soft-start process.

[0067] S3. Construct a comprehensive reward function that includes a startup result indicator factor, a peak current ratio, and a time constraint term. The comprehensive reward function is configured as follows: when the submersible electric pump starts successfully, a positive reward is calculated based on the peak current ratio using a power function; when the startup fails, a negative penalty is calculated based on the stall duration using an exponential function; and the feedback reward value for a single startup process is output in combination with the time constraint term.

[0068] Specifically, the design of the comprehensive reward function aims to guide the model to minimize current surges while ensuring successful startup. The comprehensive reward function satisfies the expression:

[0069]

[0070] In the formula, These are the weighting coefficients. As an indicator of the start-up results, Rated current, This represents the measured peak current. The current sensitivity index, For the duration of the stall, To create a penalty index for traffic jams, This is the actual startup time. For ideal startup time, This is for feedback and reward values.

[0071] Startup result indicator factor The determination method is as follows: if the motor current is detected to drop back to within a preset multiple of the rated current and the speed meets the standard, then The value is 1; if stall protection, overcurrent protection, or timeout prevents the start-up from being completed, then... The value is 0.

[0072] In actual operation, the weighting coefficient is set to Current sensitivity index Congestion penalty index .

[0073] For example, assume the rated current of the submersible electric pump Ideal startup time .

[0074] Under high-impact start-up conditions, i.e., successful start-up but excessive current, at this time... , , Substituting into the above formula, we get:

[0075] ;

[0076] Under stable startup conditions, i.e., successful startup and stable current, at this time... , , Substituting into the above formula, we get:

[0077] ;

[0078] In the case of a stalled rotor condition, i.e., a start-up failure and a stalled rotor fault, at this time... Congestion time , Substituting into the above formula, we get:

[0079] .

[0080] The comparison shows that the smooth start condition received the highest positive reward, the high-impact start condition received a very low reward or even a negative reward, and the fault-locked condition was severely penalized.

[0081] By introducing a comprehensive reward function, complex physical constraints are transformed into explicit mathematical optimization objectives, enabling neural networks to automatically balance startup reliability and grid security, thus forcing the strategy to evolve towards lower current and higher success rates.

[0082] S4. Using the deep deterministic policy gradient algorithm, the dual neural network model is iteratively trained through the experience replay mechanism and the feedback reward value; the state vector is input into the executor network after training convergence to generate the optimal soft start control parameters for the current working condition.

[0083] Specifically, an experience replay pool is established, and at each time step... It will include the current state. Execution of actions Feedback Value and the state at the next moment Quadruple The data is stored in the experience replay pool. During training, batches of data are randomly sampled from the experience replay pool.

[0084] First, the critic network parameters are updated by minimizing the loss function, which satisfies the expression:

[0085]

[0086] In the formula, The size of the sampled batch of data. For the first in the batch data Index of each sample, For the first The current state of each sample For the first The current action of each sample. For the critic's network parameters, For the critic network in parameters Next to the The current state of each sample and actions The evaluation value, This is the target value.

[0087] This expression calculates the mean squared error between the critic network's predicted scores and the actual target values. The goal of training the critic network is to minimize this mean squared error. This is to make the evaluations of the critics' network more and more accurate.

[0088] Target value Satisfying the expression:

[0089]

[0090] In the formula, For the first The feedback reward value for each sample, As a discount factor, For the first The state of each sample record at the next moment. and These are the critic target network and the actor target network, respectively. and These are the network parameters for the critic and the network parameters for the executor, respectively. For the executor target network in parameters Next to the The recommended best action to take based on the next state of each sample record.

[0091] Then, by maximizing the critic network Values ​​are used to update the executor network parameters. That is, to find a way to Gradient ascent is performed in the direction of the action with the largest value. Finally, a soft update strategy is used to update the corresponding target network parameters to maintain training stability.

[0092] In practical applications, when the submersible electric pump is about to start, the system collects data in real time to construct the current state vector, inputs it into the trained executor network to obtain the optimal action; the soft starter executes the optimal action to control the thyristor conduction angle, and calculates the feedback reward value after the start-up is completed, and stores the start-up data in the experience playback pool for subsequent incremental training.

[0093] The following combination Figure 2 The effects of the present invention will be further explained.

[0094] Figure 2The current variation trends during soft start-up of the submersible electric pump (SEP) were compared under the same downhole load conditions using the fixed parameter mode of existing technology and the neural network optimization mode of this invention. As shown in the figure, the curve of the existing technology exhibits a steep rising edge, indicating that the voltage was applied too quickly or the initial voltage setting was too high, causing the current to reach a peak of approximately 550A in a very short time. After reaching the peak, the current drops rapidly but is accompanied by some oscillations. The 550A current surge will generate a huge voltage drop on the high-impedance grid, easily triggering undervoltage protection and causing a complete platform shutdown. The curve of this invention exhibits a gentle rising edge, indicating that the actuator network outputs a longer ramp-up time and a suitable initial voltage; the peak current is precisely controlled at around 300A, approximately three times the rated current. Compared with the existing technology, the surge current is reduced by nearly 45%, and the final current drops back to approximately 100A, indicating that the motor did not stall. This proves that this invention provides sufficient starting torque while significantly reducing the current.

[0095] Thus, by using the offline training and experience replay mechanism of the deep deterministic policy gradient algorithm, the correlation of time series data is broken, and the model achieves stable convergence. At the same time, the continuous learning mechanism during online application enables the system to self-correct using each startup data, and has the ability to evolve throughout the entire life cycle to adapt to equipment aging and environmental changes.

[0096] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for optimizing the parameters of a submersible electric pump soft starter based on neural networks, characterized in that, Includes the following steps: Acquire downhole environmental characteristic data, historical state characteristic data and power grid environmental characteristic data of submersible electric pumps, and construct a state vector representing the current operating condition after normalizing the acquired data; A dual neural network model based on an actor-critic architecture is constructed; wherein, the actor network is used to output the corresponding action policy according to the state vector, and the action policy includes soft-start control parameters; the critic network is used to evaluate the expected value of the action policy; A comprehensive reward function is constructed, comprising a startup result indicator factor, a peak current ratio, and a time constraint term. The comprehensive reward function is configured as follows: when the submersible electric pump starts successfully, a positive reward is calculated using a power function based on the peak current ratio; when startup fails, a negative penalty is calculated using an exponential function based on the stall duration; and the feedback reward value for a single startup process is output in conjunction with the time constraint term. The comprehensive reward function satisfies: These are the weighting coefficients. As an indicator of the start-up results, Rated current, This represents the measured peak current. The current sensitivity index, For the duration of the stall, To create a penalty index for traffic jams, This is the actual startup time. For ideal startup time, For feedback and reward value; The dual neural network model is iteratively trained using a deep deterministic policy gradient algorithm, through an experience replay mechanism and the feedback reward value; the state vector is input into the executor network after training convergence to generate the optimal soft-start control parameters for the current operating condition.

2. The method for optimizing the parameters of a submersible electric pump soft starter cabinet based on a neural network according to claim 1, characterized in that, The downhole environmental characteristic data includes bottom hole fluid temperature data and bottom hole fluid pressure data; the historical state characteristic data includes current downtime data; and the power grid environmental characteristic data includes real-time power grid bus voltage data and power grid frequency data. The state vector satisfies the expression: In the formula, For normalization function, The temperature of the fluid at the bottom of the well. The bottom hole fluid pressure, This is the current downtime. This represents the real-time voltage of the power grid bus. This refers to the power grid frequency.

3. The method for optimizing the parameters of a submersible electric pump soft starter based on a neural network according to claim 2, characterized in that, The soft-start control parameters include initial voltage percentage, ramp-up time, and current limiting setting; the output layer of the actuator network adopts a hyperbolic tangent activation function, and the output result is mapped to the device safety threshold range through linear transformation; The safety threshold range of the device is: initial voltage percentage of 30% to 80%, ramp-up time of 5s to 60s, and current limiting setting of 2.0 to 5.0 times the rated current.

4. The method for optimizing the parameters of a submersible electric pump soft starter cabinet based on a neural network according to claim 2, characterized in that, The startup result indicator factor The determination method is as follows: if the motor current is detected to drop back to within a preset multiple of the rated current and the speed meets the standard, then The value is 1; If stall protection, overcurrent protection, or timeout prevents startup from being completed, then The value is 0.

5. The method for optimizing the parameters of a submersible electric pump soft starter cabinet based on a neural network according to claim 1, characterized in that, The iterative training process of the dual neural network model includes: updating the critic network parameters by minimizing the loss function, and minimizing the loss function. Satisfying the expression: In the formula, The size of the sampled batch of data. For the first in the batch data Index of each sample, For the first The current state of each sample For the first The current action of each sample. For the critic's network parameters, For the critic network in parameters Next to the The current state of each sample and actions The evaluation value, This is the target value.

6. The method for optimizing the parameters of a submersible electric pump soft starter cabinet based on a neural network according to claim 5, characterized in that, The target value Satisfying the expression: In the formula, For the first The feedback reward value for each sample, As a discount factor, For the first The state of each sample record at the next moment. and These are the critic target network and the actor target network, respectively. and These are the network parameters for the critic and the network parameters for the executor, respectively. For the executor target network in parameters Next to the The recommended best action to take based on the next state of each sample record.

7. The method for optimizing the parameters of a submersible electric pump soft starter cabinet based on a neural network according to claim 1, characterized in that, The gradient algorithm utilizing a deep deterministic strategy includes the following steps: Establish an experience replay pool. At each time step, store a quadruple sample containing the current state, the action performed, the feedback value, and the state at the next time step into the experience replay pool. Batch data is randomly sampled from the experience replay pool to update the network parameters of the executor network and the critic network, and then the corresponding target network parameters are updated using a soft update method.

8. The method for optimizing the parameters of a submersible electric pump soft starter cabinet based on a neural network according to claim 7, characterized in that, The state vector is input into the converged executor network to generate optimal soft-start control parameters for the current operating condition, including: When the submersible electric pump is about to start, real-time data is collected to construct the current state vector, which is then input into the trained executor network to obtain the optimal action. The soft starter executes the optimal action to control the thyristor conduction angle, and calculates the feedback value after startup. The startup data is then stored in the experience playback pool for subsequent incremental training.

9. The method for optimizing the parameters of a submersible electric pump soft starter cabinet based on a neural network according to claim 1, characterized in that, Both the executor network and the critic network contain two fully connected hidden layers; the first fully connected hidden layer has 256 nodes and the second fully connected hidden layer has 128 nodes, and both hidden layers use the ReLU activation function.

Images