A diesel generator set based on preheating time prediction and control method thereof

Abstract

The present invention relates to the field of diesel generator set control technology, and discloses a diesel generator set based on preheating time prediction and a control method thereof, comprising the following steps: step S101, obtaining the working information of the engine unit; step S102, constructing a preheating time prediction model, inputting the working information of the engine unit into the preheating time prediction model, and outputting the predicted preheating time of the engine unit; step S103, controlling the preheating time of the engine unit and the maximum load of the generator set. The present invention comprehensively considers the influence of the use time, maintenance, air quality, temperature and humidity of the diesel generator set on the preheating time and performance of the diesel generator set, and can determine the preheating time and maximum safe load of the diesel generator set before using it, which can avoid the phenomenon that the diesel generator set fails due to insufficient understanding of the current status of the diesel generator set during manual operation, and effectively improves the service life of the diesel generator set.

Description

Technical Field

[0001] The present invention relates to the technical field of diesel generator control, and more specifically, to a diesel generator set based on preheating time prediction and a control method thereof. Background Art

[0002] When a diesel generator set is just started, the internal temperature is low, the diesel combustion is not complete, the energy conversion efficiency is low, and due to the characteristics of the material, the oil has not reached the appropriate temperature at the beginning of operation, and the temperatures of components such as the piston and cylinder may vary greatly. If it runs at high load at this time, the diesel generator set will experience a shortened service life and failures.

[0003] In addition, due to the extended working hours of the diesel generator set and environmental factors, its performance will gradually decline. If it operates beyond its maximum performance, it will cause damage to the diesel generator set.

[0004] A common solution is to employ skilled technical workers to control the preheating of the diesel engine unit and the maximum load of the generator unit. However, it is impossible for humans to perform precise control according to the actual situation of the diesel generator set, which will inevitably lead to damage or even destruction of the diesel generator set. Therefore, a high-performance diesel generator set and a control method thereof are proposed. Summary of the invention

[0005] The present invention provides a diesel generator set based on preheating time prediction and a control method thereof, which solve the technical problems in the above-mentioned background technology.

[0006] The present invention provides a diesel generator set based on preheating time prediction, comprising an engine unit, a generator unit, a cooling system, a fuel system, a control system, a silencer system, an exhaust system, and a shock absorption system; the engine is started and stopped, data is measured, displayed, and fault protection is performed through a controller, and the power quantity of the generator unit is measured, displayed, and protected; the controller is also used to predict and control the safe preheating time of the engine unit, and then control the maximum load of the generator unit.

[0007] A control method for a diesel generator set based on preheating time prediction of the present invention comprises the following steps:

[0008] Step S101, obtaining the working information of the generator set;

[0009] The working information of the generator set includes: the generator usage time parameter, load level parameter, maintenance status parameter, fuel quality parameter, temperature and humidity parameter and air quality parameter;

[0010] Step S102, constructing a preheating time prediction model, inputting the working information of the engine unit into the preheating time prediction model, and outputting the predicted preheating time of the engine unit;

[0011] Step S103, calculating the performance degradation index of the generator set according to the above parameters, and controlling the preheating time and the maximum load of the engine unit according to the predicted preheating time of the engine unit and the performance degradation index of the generator set.

[0012] Further, the method for obtaining the working information of the generator set includes:

[0013] The factory time and actual recorded usage time of the generator set are used as the usage time parameter of the generator set;

[0014] Obtain the maximum output power and average output power of the generator set in the latest month as load level parameters;

[0015] Obtain the maintenance times of the generator set and calculate the maintenance status parameters based on the maintenance times. The calculation formula is as follows:

[0016] Maintenance status parameter = usage time parameter / maintenance times;

[0017] The fuel purity, water content and viscosity are tested by testing equipment, and the fuel quality parameters are calculated based on the fuel purity, water content and viscosity. The calculation formula is as follows:

[0018] Fuel quality parameters = a*fuel purity + b*fuel water content + c*fuel viscosity;

[0019] Where a, b and c represent the first, second and third weight factors respectively; a, b and c are all custom parameters;

[0020] The temperature and humidity information of the area where the generator set is located is obtained as temperature and humidity parameters. The temperature and humidity information is obtained through a thermometer and a hygrometer;

[0021] The air quality information published by the environmental monitoring station is obtained as the air quality parameter.

[0022] Further, the warm-up duration prediction model includes a first hidden layer and a classifier;

[0023] The working information of the engine unit is input into the first hidden layer, and the first updated feature is output;

[0024] The first updated feature is input into the classifier, and the classifier outputs the predicted preheating time of the engine unit, where the predicted preheating time represents the safe preheating time of the engine unit.

[0025] Furthermore, the first hidden layer of the warm-up duration prediction model is constructed based on a multi-layer perceptron.

[0026] Further, the working information of the m engine units is obtained as training samples for training the preheating time prediction model;

[0027] Obtain the time when the intake air temperature of each engine unit reaches the preset temperature as the sample label for training the preheating time prediction model;

[0028] m is a custom parameter.

[0029] Furthermore, the calculation formula of the warm-up time prediction model includes:

[0030] The loss function of the warm-up time prediction model is as follows:

[0031]

[0032] Where L represents the mean square error loss of the warm-up time prediction model; n represents the number of training samples; y i represents the predicted preheating time of the i-th engine unit output by the preheating time prediction model; represents the actual preheating time of the i-th engine unit;

[0033] The gradient descent formula of the loss function of the warm-up time prediction model is as follows:

[0034]

[0035] In the formula, w j1 represents the updated weight parameter in the warm-up time prediction model; w j represents the weight parameter before updating in the warm-up time prediction model; α represents the learning rate of the warm-up time prediction model; The loss function of the warm-up time prediction model is w j The partial derivative of .

[0036] Further, the calculation formula for calculating the performance degradation index of the generator set includes:

[0037] PDI=β1*ΔT+β2*M f +β3*F q +β4*E f ;

[0038] Where PDI represents the performance degradation index of the generator set; ΔT represents the difference between the predicted preheating time of the engine unit output by the preheating time prediction model and the average preheating time of the newly produced engine unit; M f Indicates the maintenance status parameter of the generator set; F q Indicates the fuel quality parameters of the generator set; E fRepresents the air quality parameter of the generator set; β1, β2, β3 and β4 represent the first, second, third and fourth weight values ​​respectively; β1, β2, β3 and β4 are all custom parameters.

[0039] Further, controlling the preheating time of the engine unit and the maximum load of the generator unit includes the following steps:

[0040] Step S201, setting the preheating time of the engine unit according to the predicted preheating time of the engine unit output by the preheating time prediction model;

[0041] Step S202, calculating the ratio of the performance degradation index of the generator set to the performance degradation index of a newly manufactured generator set;

[0042] Step S203, according to the calculation result of step S202, the maximum load of the generator unit is set to a first load value, the first load value = the maximum load of the new generator * the ratio of the performance degradation index of the generator set to the performance degradation index of the new generator set.

[0043] The beneficial effects of the present invention are as follows: it comprehensively considers the influence of the use time, maintenance, air quality, temperature and humidity of the diesel generator set and other information on the preheating time and performance of the diesel generator set, and can determine the preheating time and maximum safe load of the diesel generator set before it is used, which can avoid the phenomenon that manual operation due to insufficient understanding of the current condition of the diesel generator set may cause the diesel generator set to malfunction, and effectively improves the service life of the diesel generator set.

[0044] Accurate prediction and optimized load management: By combining the model predictive control (MPC) algorithm, historical data and real-time monitoring data can be used to accurately predict the future state and performance changes of the generator set. The MPC algorithm provides high-precision load forecasting through the time series prediction ARIMA model to ensure that the generator set can maintain the best state in future operation, and continuously optimize the maximum load setting with the help of the deep Q network (DQN). DQN dynamically adjusts the load setting of the generator unit by learning the complex relationship between historical and real-time data to improve overall performance.

[0045] Real-time response and dynamic adjustment: By installing sensors in the generator set to monitor key performance indicators such as temperature, pressure, and vibration in real time, and inputting these data into the neural network model, the model can accurately reflect the current status. This real-time monitoring and feedback mechanism enables the generator set to respond quickly to emergencies and ensure safety and stability. According to the changes in real-time monitoring data, the input of the neural network model is dynamically adjusted. The fusion of real-time monitoring data and historical data can enhance the model's predictive ability and improve the accuracy and response speed of load management.

[0046] Improved safety and reliability: When dynamically adjusting the load settings, safety constraints are introduced to ensure that the load settings do not cause overload or other safety issues of the generator unit. By setting parameters such as maximum temperature and pressure, we ensure the safe operation of the generator unit while optimizing the load. We designed a reward function that comprehensively considers power output, temperature, pressure, and vibration. This reward function not only guides the model to optimize the load settings, but also ensures the stability and safety of the generator unit under various operating conditions.

[0047] Adaptive learning and optimization: The deep reinforcement learning model has the ability of adaptive learning. By continuously obtaining feedback from the environment and updating strategies, the model can still maintain efficient operation under changing working conditions. Combining model predictive control and deep reinforcement learning, the intelligent algorithm can continuously optimize itself during operation and improve its ability to handle complex nonlinear relationships, thereby achieving optimal load management of the generator unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the overall flow chart of the present invention;

[0049] Figure 2 It is a flow chart for controlling the preheating time of the engine unit and the maximum load of the generator unit of the present invention.

[0050] Figure 3 This is the flow chart after adding intelligent algorithm DETAILED DESCRIPTION

[0051] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0052] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in one or more embodiments of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0053] Embodiment 1 of the present invention provides a diesel generator set based on preheating time prediction and a control method thereof, comprising an engine unit, a generator unit, a cooling system, a fuel system, a control system, a silencing system, an exhaust system, and a shock absorbing system;

[0054] The controller is used to predict and control the safe preheating time of the engine unit and then control the maximum load of the generator unit.

[0055] The second embodiment of the present invention is that Figure 1-2 As shown, a high-performance diesel generator set is provided, which includes the following steps:

[0056] Step S101, obtaining the working information of the generator set;

[0057] The working information of the generator set includes: the generator usage time parameter, load level parameter, maintenance status parameter, fuel quality parameter, temperature and humidity parameter and air quality parameter;

[0058] Step S102, constructing a preheating time prediction model, inputting the working information of the engine unit into the preheating time prediction model, and outputting the predicted preheating time of the engine unit;

[0059] Step S103, calculating the performance degradation index of the generator set according to the above parameters, and controlling the preheating time and the maximum load of the engine unit according to the predicted preheating time of the engine unit and the performance degradation index of the generator set.

[0060] In one embodiment of the present invention, a method for obtaining operating information of a generator unit includes:

[0061] The factory time of the generator set and the actual recorded usage time are used as the usage time parameters of the generator set. The factory time of the generator set indicates the time when the diesel generator set is produced. If it is not used for a long time after production, its conversion efficiency will also decrease to a certain extent due to material aging, material oxidation and oil deposition. The actual recorded usage time indicates the actual working time of the diesel generator set. During the working process of the diesel generator set, the wear between the piston and the cylinder, the wear between the various components and other factors will also affect the conversion efficiency of the diesel generator set.

[0062] The maximum output power and average output power of the generator unit in the last month are obtained as load level parameters. By testing the output current and voltage of the diesel generator, its output power can be calculated, and then the maximum output power is selected. In addition, the load with the longest working time of the diesel generator power is selected as the average output power;

[0063] Obtain the maintenance times of the generator set and calculate the maintenance status parameters according to the maintenance times. The calculation formula is as follows:

[0064] Maintenance status parameter = usage time parameter / maintenance times;

[0065] The fuel purity, water content and viscosity are tested by testing equipment, and the fuel quality parameters are calculated based on the fuel purity, water content and viscosity. The calculation formula is as follows:

[0066] Fuel quality parameters = a*fuel purity + b*fuel water content + c*fuel viscosity;

[0067] Wherein, a, b and c represent the first, second and third weight factors respectively, a, b and c are all custom parameters, and a+b+c=1;

[0068] The purity, water content and viscosity of the fuel are obtained by testing with a gas chromatograph, a portable water content detector and a portable rapid viscosity detector, respectively. The gas chromatograph, the portable water content detector and the portable rapid viscosity detector are all prior arts and will not be described in detail here.

[0069] The temperature and humidity information of the area where the generator set is located is obtained as temperature and humidity parameters. The temperature and humidity information is obtained through a thermometer and a hygrometer;

[0070] The air quality information published by the environmental monitoring station is obtained as the air quality parameter.

[0071] In one embodiment of the present invention, the preheating duration prediction model includes a first hidden layer and a classifier;

[0072] The working information of the engine unit is input into the first hidden layer, and the first updated feature is output;

[0073] The first updated feature is input into the classifier, and the classifier outputs the predicted preheating time of the engine unit, where the predicted preheating time represents the safe preheating time of the engine unit.

[0074] In one embodiment of the present invention, the first hidden layer of the preheating duration prediction model is constructed based on a multi-layer perceptron.

[0075] In one embodiment of the present invention, the operating information of m engine units is obtained as training samples for training a preheating time prediction model;

[0076] Obtain the time when the intake air temperature of each engine unit reaches the preset temperature as the sample label for training the preheating time prediction model;

[0077] m is a custom parameter, and m is preferably 2000.

[0078] In one embodiment of the present invention, the calculation formula of the preheating time prediction model includes:

[0079] The loss function of the warm-up time prediction model is as follows:

[0080]

[0081] Where L represents the mean square error loss of the warm-up time prediction model; n represents the number of training samples; y i represents the predicted preheating time of the i-th engine unit output by the preheating time prediction model; represents the actual preheating time of the i-th engine unit;

[0082] The gradient descent formula of the loss function of the warm-up time prediction model is as follows:

[0083]

[0084] In the formula, w j1 represents the updated weight parameter in the warm-up time prediction model; w j represents the weight parameter before updating in the warm-up time prediction model; α represents the learning rate of the warm-up time prediction model; The loss function of the warm-up time prediction model is w j The partial derivative of .

[0085] In one embodiment of the present invention, the calculation formula for calculating the performance degradation index of the generator set includes:

[0086] PDI=β1*ΔT+β2*M f +β3*F q +β4*E f ;

[0087] Where PDI represents the performance degradation index of the generator set; ΔT represents the difference between the predicted preheating time of the engine unit output by the preheating time prediction model and the average preheating time of the newly produced engine unit; M f Indicates the maintenance status parameter of the generator set; F q Indicates the fuel quality parameters of the generator set; E f represents the air quality parameter of the generator set; β1, β2, β3 and β4 represent the first, second, third and fourth weight values ​​respectively; β1, β2, β3 and β4 are all custom parameters, and β1+β2+β3+β4=1.

[0088] In one embodiment of the present invention, controlling the preheating time of the engine unit and the maximum load of the generator unit includes the following steps:

[0089] Step S201, setting the preheating time of the engine unit according to the predicted preheating time of the engine unit output by the preheating time prediction model;

[0090] Step S202, calculating the ratio of the performance degradation index of the generator set to the performance degradation index of a newly manufactured generator set;

[0091] Step S203, according to the calculation result of step S202, the maximum load of the generator unit is set to a first load value, the first load value = the maximum load of the new generator * the ratio of the performance degradation index of the generator set to the performance degradation index of the new generator set.

[0092] In steps S201-S203, when controlling the preheating time of the engine unit and the maximum load of the generator unit, deep reinforcement learning and model predictive control intelligent algorithms are introduced to optimize the control process of the preheating time of the engine unit and the maximum load of the generator unit. The algorithm enhances accurate prediction and optimized load management, and optimizes real-time response and dynamic adjustment. The specific process is as follows:

[0093] The specific process of step S201 is:

[0094] Step 1: First, collect historical data of the generator set startup for training the model; these historical data include startup time, ambient temperature, humidity, and load size; represent these data as a set:

[0095]

[0096] in:

[0097] ti represents the start time of the i-th generator set;

[0098] Ti represents the corresponding ambient temperature;

[0099] Hi indicates the ambient humidity;

[0100] Li represents the load size;

[0101] These data will serve as input features for training the model;

[0102] Step 2: In order to simulate the actual application scenario, the collected historical data set D is divided into a training data set Dtrain and a test data set Dtest in chronological order;

[0103] Step 3: Neural network model construction, select convolutional neural network CNN as the model to solve the engine unit preheating time prediction problem;

[0104] The CNN model of the network structure design includes convolutional layer, pooling layer and fully connected layer; each layer is described in detail below;

[0105] Convolutional layer: The convolutional layer is used to extract features from the input data. Multiple convolution kernels will be used to capture the features of high-performance diesel generators at different time and space scales. There are Nc convolution kernels, each of size K×K, then the output of the convolutional layer can be expressed as:

[0106] O c =ReLU(W c *X+b c )

[0107] Among them, O c is the output of the convolutional layer, W c is the weight of the convolution kernel, X is the input data, b c is the bias term, * indicates the convolution operation, and ReLU is the activation function;

[0108] Known step 1 training data set D train The size of input data X is W in ×H in ×C in , where W in Indicates width, H in Indicates height, C in Indicates the number of channels;

[0109] Convolution kernel W c The size is K×K×C in ×N c , where K represents the width and height of the convolution kernel, N c Represents the number of convolution kernels, convolution operation W c *X performs a convolution operation on the convolution kernel and the input data, and the resulting feature map size is W out ×H out ×N c, where W out and H out Represent the width and height of the feature map respectively, depending on the size and step size of the convolution kernel, and then add the result of the convolution operation to the bias term b c Add them together and perform nonlinear transformation through the ReLU activation function to get the final convolutional layer output;

[0110] Pooling layer: The pooling layer is used to reduce the dimension of the feature map while retaining the most important features; the maximum pooling operation is selected; the size of the pooling layer is F×F, then the output of the pooling layer can be expressed as:

[0111] O p =MaxPool(O c )

[0112] Among them, O p is the output of the pooling layer, and MaxPool represents the maximum pooling operation;

[0113] Known step 1 training data set D train The size of input data X is W in ×H in ×C in , where W in Indicates width, H in Indicates height, C in Indicates the number of channels,

[0114] The size of the pooling layer is F×F, which represents the size of each pooling window. For each pooling window, the maximum pooling operation will select the maximum value in the window as the output. The step size of the pooling operation is S, so the size of the feature map after pooling is:

[0115]

[0116] Among them, W out and H out Respectively represent the width and height of the feature map after pooling;

[0117] For each pooling window, the max pooling operation selects the maximum value within the window as the output; therefore, the output of the pooling layer can be expressed as:

[0118] O p (i,j,k)=max(O c (i′S:i′S+F,j′S:j′S+F,k))

[0119] Among them, O p (i, j, k) represents the value of the i-th row, j-th column, and k-th channel output by the pooling layer. c(i′S:i′S+F,j′S:j′S+F,k) represents the area corresponding to the pooling window in the input feature map, and k represents the number of channels of the input feature map;

[0120] Finally, the output of the pooling layer is obtained p , where the maximum value of each pooling window in the input feature map is retained;

[0121] The fully connected layer is used to connect the outputs of the convolutional layer and the pooling layer and generate the final prediction result; the output of the fully connected layer is directly used as the predicted warm-up duration; the output size of the fully connected layer is M, then the output of the fully connected layer can be expressed as:

[0122] O f =Softmax(W f O p +b f )

[0123] Among them, O f is the output of the fully connected layer, W f is the weight of the fully connected layer, b f is the bias term, Softmax is the activation function;

[0124] The output of the pooling layer is O p The size of is W×H×C, where W represents width, H represents height, and C represents the number of channels. The output of the pooling layer is flattened into a vector of size N, where N = W×H×C;

[0125] The weight of the fully connected layer W f The size of is N×M, where M represents the output size of the fully connected layer;

[0126] The output of the fully connected layer can be expressed as the result of matrix multiplication:

[0127] O f =W f O p +b f

[0128] Among them, b f is the bias term of the fully connected layer, with dimension M;

[0129] The output of the fully connected layer is processed by the activation function Softmax to convert it into a probability distribution; the definition of the Softmax function is as follows:

[0130]

[0131] Among them, z i represents the i-th element of the input vector, and M represents the dimension of the vector;

[0132] Finally, the output O of the fully connected layer is obtained f , each element of which represents the probability of the corresponding category;

[0133] After defining the network structure, the training data set D train Input into the model for training; during the training process, continuously adjust the model parameters to make the prediction results as close as possible to the actual warm-up time; finally, use the test data set to evaluate the performance of the model and check its generalization ability on new data;

[0134] Step 4: Deep reinforcement learning training. The deep Q network DQN is an algorithm that combines deep learning and reinforcement learning. In this step, the DQN algorithm is selected to train the convolutional neural network CNN model, and CNN is used to approximate the Q value function to solve the problem of engine unit preheating time prediction;

[0135] Design a reward function to guide the model to learn the appropriate warm-up time, so that the model can get a higher reward when the predicted warm-up time is close to the actual time, so as to gradually approach the optimal prediction through repeated training; the reward function R(s,a) is defined as the negative value of the difference between the actual warm-up time and the expected warm-up time, that is:

[0136] R(s,a)=-|T actual -T expected |

[0137] in:

[0138] T actual is the actual preheating time;

[0139] T expected Expected warm-up time;

[0140] Initialize the weights and bias items of the convolutional neural network: The weight θ of the convolution kernel and the bias item b are randomly initialized; for the initialization of the convolution layer, pooling layer and fully connected layer, the Xavier initialization method is used;

[0141] Construct a DQN framework that includes a target network and an online network; the online network is used to generate Q values, and the target network is updated after a certain number of steps to provide a more stable target value. The parameters of the target network are θ - Copy from the online network parameters θ and update every fixed number of steps to prevent oscillation during training;

[0142] The experience replay pool is used to store experience samples (s, a, r, s′) generated by the interaction between the model and the environment, namely the current state, action, reward, and next state; these samples will be randomly extracted during the training process for model update.

[0143] Before starting deep reinforcement training, the current state s is obtained from the environment, that is, the current start-up time, ambient temperature, humidity and load size of the generator set; these features constitute the state space, and then the action a is selected according to the ∈-greedy strategy, that is, the predicted warm-up time; the exploration rate ∈ determines whether the model chooses a random action or the optimal action according to the current strategy; as the training progresses, ∈ gradually decreases; when the optimal action is selected, the action a is executed, and the new state s′ and reward r are observed; the new state s′ is the new feature of the generator after executing action a, and the reward r is calculated by the reward function R(s,a); then the experience samples (s,a,r,s′) are stored in the experience replay pool; these samples will be used for subsequent training;

[0144] Randomly draw a small batch of samples (s i ,a i ,r i ,s′ i ); These samples are used to calculate the target Q value and update the parameters of the online network;

[0145] Calculate the target Q value:

[0146] y i =r i +γmax a′ Q(s′ i ,a′;θ - )

[0147] in:

[0148] γ is a discount factor that weighs the impact of immediate rewards and future rewards;

[0149] θ- is the parameter of the target network,

[0150] maxa ′ Q(si′,a′;θ-) represents the Q value of selecting the optimal action in the new state si′;

[0151] Calculate the loss function:

[0152]

[0153] E is the dependent variable of the loss function. The loss function measures the gap between the target Q value and the current Q value. The parameter θ of the online network is optimized by minimizing the loss function.

[0154] Update the parameters θ of the online network through the gradient descent method; calculate the gradient and update the parameters through the back propagation algorithm, and copy the parameters of the online network to the target network θ- at fixed steps;

[0155] Q value update formula:

[0156]

[0157] in:

[0158] α is the learning rate, which controls the step size of parameter update;

[0159] γ is a discount factor that weighs the impact of immediate rewards and future rewards;

[0160] θ is the parameter of the online network;

[0161] θ- are the parameters of the target network;

[0162] Parameter variable explanation

[0163] State s: the current start time of the generator set, ambient temperature, humidity, load size and other characteristics; these characteristics serve as inputs to the state space and constitute a description of the environment;

[0164] Action a: predicted warm-up duration; each action in the action space corresponds to a predicted value of warm-up duration;

[0165] Reward r: calculated by the reward function R(s,a), used to evaluate the difference between the predicted warm-up time and the actual warm-up time;

[0166] Experience samples ((s,a,r,s')): including the current state, action, reward and next state. These samples are used for model training to improve the prediction performance of the model.

[0167] Target Q value y i : In the new state s′ i The Q value of the optimal action is selected to calculate the loss function and optimize the parameters of the online network;

[0168] Discount factor γ: used to weigh the impact of immediate rewards and future rewards, usually between 0 and 1;

[0169] Learning rate α: controls the step size of parameter updates and affects the convergence speed and stability of the model;

[0170] Step 5: Predict the preheating time. Before the generator set is started, the preheating time needs to be predicted based on the current environmental conditions and historical data to ensure that the generator set reaches the best working state when it starts. By inputting these input data into the trained CNN model, the model can output the optimal preheating time.

[0171] Input data: Current environmental conditions: Ambient temperature T env 、Humidity env , Current startup time t start , load size; historical data: historical warm-up time T history , start time thistory , environmental conditions records;

[0172] The above data are preprocessed and the input data is standardized to ensure that they have similar scales to avoid certain features having a greater impact on model training. The historical data is processed with a sliding window to extract time series features, the conditions of the past few startups, and the trends of environmental conditions; they are input into the trained CNN model, which extracts and learns features from the input data through a series of convolutional layers, pooling layers, and fully connected layers;

[0173] Model output prediction warm-up time T predicted , represents the optimal preheating time predicted based on the current environment and historical data, and the model predicted value T predicted , adjust the preheating time of the engine unit: if T predicted Larger: Increase the preheating time to ensure that the engine reaches the optimal operating temperature and prevent the engine from being damaged or causing efficiency loss due to insufficient temperature during startup. predicted Smaller: Reduce preheating time, save energy, and avoid unnecessary long preheating time;

[0174] The specific process of step S202 is:

[0175] Step 1: Collect and process the historical operating data of the generator set and extract some key features to facilitate the subsequent calculation of the performance degradation index and formulation of the optimization strategy;

[0176] To facilitate the subsequent calculation of performance degradation index and formulation of optimization strategies;

[0177] Usage time usage_time: For the generator set, the usage time is obtained by monitoring the start and stop time of the equipment to record the usage time of the generator set, including the total running time and the running time of each start;

[0178] Load history load_history: monitors load data in real time by connecting sensors or monitoring systems, and records the load conditions of the generator unit in different time periods, including the size and changes of the load;

[0179] Maintenance record maintenance_record: Information is recorded in the maintenance log or repair record, which can be collected through manual records or maintenance management systems; records the maintenance history of the generator set, including the time, content and frequency of maintenance;

[0180] Running time running_time: calculate the total running time of the generator set, which can be calculated by the usage time usage_time;

[0181] Load variation trend load_variation_trend: Use time series analysis method to analyze load history load_history, extract load variation trend, load fluctuation frequency and amplitude;

[0182] Step 2: Build a deep Q network DQN deep reinforcement learning model and train it with historical data to obtain an accurate performance degradation index;

[0183] Deep reinforcement learning model construction: The state space state takes the key features of the generator set as the state, the running time running_time and the load change trend load_variation_trend as the input of the deep Q network; the action space action: represents the load setting of the generator unit;

[0184] The output is a performance degradation index, which indicates the degree of performance degradation of the generator set;

[0185] Design deep neural network structures, including convolutional layers, pooling layers, and fully connected layers, as well as corresponding activation functions and loss functions;

[0186] Use the historical operating data of the generator sets collected previously, including usage time, load history, and maintenance records, as training data for the model; input the historical data into the deep reinforcement learning model, and use the back-propagation algorithm and Adam optimizer to update the parameters of the model to minimize the loss function;

[0187] The mean square error (MSE) loss function is used in the present invention to update the parameters of the model by optimizing the loss;

[0188] The specific loss function can be expressed as:

[0189] Loss=MSE(Q(state,action),Qtarget)

[0190] Where Qtarget is the target Q value, calculated by the Bellman equation;

[0191] According to the back-propagation algorithm and Adam optimizer, the parameters of the deep Q network are updated so that the model can better fit the training data and reward function;

[0192] Define a reward function so that the model can learn the appropriate performance degradation index; the reward function should take into account the performance changes of the generator set, giving negative rewards when the performance degrades significantly, and positive rewards when the performance is stable or improved;

[0193] Design the reward function R so that the model can obtain the appropriate reward value according to the current state and action; the specific reward function can be expressed as:

[0194] R(state,action)=f(state,action)

[0195] Performance evaluation in practical applications: Input the current generator set status into the trained deep reinforcement learning model to obtain the corresponding performance degradation index;

[0196] According to the calculated performance degradation index, corresponding optimization measures are taken to adjust the load setting and maintenance frequency of the generator unit to minimize the performance degradation; this provides an accurate basis for the maximum load setting in the subsequent step S203;

[0197] The dynamic adjustment process of the maximum load in step S203 is:

[0198] Step 1: The model predictive control algorithm uses historical data and current status to predict the future status and performance changes of the generator unit, and dynamically adjusts the maximum load based on the prediction results to enable the generator unit to achieve the best working state while ensuring safety;

[0199] Design a model predictive control algorithm and use various time series prediction methods to predict the future state and performance changes of the generator set; select the ARIMA model, whose basic formula is:

[0200] Y t =c+φ1Y t-1 +φ2Y t-2 +...+φ p Y t-p +∈ t +θ1∈ t-1 +θ2∈ t-2 +...+ θ q∈ t-q

[0201] Among them, Y t is the observed value of the time series, c is a constant, Φ1φ2...Φ p is the autoregressive coefficient θ1θ2...θ q is the moving average coefficient, ∈ t ∈ t-1 ...∈ t-q is the white noise error term;

[0202] Combined with deep reinforcement learning algorithms, a neural network model is constructed that can learn and optimize the maximum load setting; the model will learn the optimal maximum load setting strategy based on the current generator unit state and the predicted future state to maximize the performance of the generator unit;

[0203] A deep Q network DQN model is used to learn and optimize the maximum load setting; its update rule has been detailed in the previous steps, where the update formula of the Q value is as follows:

[0204] Q(s,a)=Q(s,a)+α(R(s,a)+γmax′ a Q(s′,a′)-Q(s,a))

[0205] Among them, Q(s,a) is the Q value of taking action a in state s, α is the learning rate, which controls the step size of parameter update, R(s,a) is the reward function, which represents the reward obtained after taking action a in state s, γ is the discount factor, which weighs the impact of immediate rewards and future rewards, and max′ a Q(s′,a′) is the Q value of choosing the best action in the new state s′;

[0206] Step 2: In the load optimization process, the recommended maximum load setting is obtained by inputting the current generator unit status into the neural network model; then, according to the safety requirements, the maximum load is dynamically adjusted to ensure that the generator unit operates in the optimal working state;

[0207] Input the current state into the neural network model, which is trained by a deep reinforcement learning algorithm. By learning historical data and the current state, it can generate the optimal maximum load setting.

[0208] The current generator unit state is represented as a vector S, which contains various performance indicators, temperature, pressure, and vibration; then, this state vector is used as the input of the neural network model;

[0209] The neural network model processes the current state and outputs a recommended maximum load setting value L reconmmended ;

[0210] Before applying the recommended maximum load setting, you need to ensure that it will not cause overloading of the generator unit or other safety issues; set the maximum temperature T max and maximum pressure P max As a security constraint;

[0211] If the recommended maximum load setting L reconmmended If the safety requirements are met, it is applied to the generator unit, that is, L adjusted =L reconmmended ; Otherwise, it is necessary to further adjust the maximum load or take other measures to ensure safety;

[0212] The process of dynamically adjusting the maximum load can be expressed by the following formula:

[0213]

[0214] Among them, L adjusted is the adjusted maximum load setting, L safe is the maximum load setting that meets safety requirements;

[0215] Step 3: In the performance monitoring phase, the operating status and performance indicators, temperature, pressure, and vibration of the generator set are monitored in real time, and the input of the neural network model is adjusted according to the real-time monitoring data to ensure that the model can make accurate predictions about the current status; this can achieve timely feedback and adjustments to cope with changes and emergencies in the operation of the generator set;

[0216] Install sensors in the generator set to collect data on various performance indicators in real time. Temperature, pressure, and vibration sensors will continuously monitor the operating status of the generator set and send the real-time collected data to the monitoring system;

[0217] The real-time monitoring data is used as one of the inputs of the neural network model so that the model can perceive the current operating status. The neural network model adopts the convolutional neural network CNN structure, and its input includes current status data and historical data. Therefore, the real-time monitoring data will constitute the input of the neural network model together with the historical data. The neural network model uses historical data and current status to predict the future status and performance changes of the generator set. In order to ensure that the model can make accurate predictions about the current status, we need to dynamically adjust the input of the neural network model according to the real-time monitoring data.

[0218] Real-time monitoring data is X real_time , historical data is X historical , then the input of the neural network model can be expressed as X = [X real_time ,X historical ];

[0219] The adjustment of real-time monitoring data mainly includes the following steps:

[0220] Preprocess the real-time monitoring data, including data cleaning, denoising, and normalization, to improve the quality and reliability of the data; the preprocessing process can be expressed as:

[0221] X preprocess =Preprocess(X real_time )

[0222] The pre-processed real-time monitoring data is integrated with the historical data to form the complete input of the neural network model. The purpose of this is to use the information of the historical data to assist the model in making predictions and improve the generalization ability of the model.

[0223] X=[Xpreprocess ,X historical ]

[0224] Dynamic input adjustment: Dynamically adjust the input of the neural network model according to the changes in real-time monitoring data; if there are sudden changes or abnormalities in the real-time monitoring data, adjust its weight higher to ensure that the model can more accurately capture the characteristics of the current state;

[0225] Enter the dynamic adjustment formula

[0226] The dynamic weight coefficient is λ, and the input of the neural network model after dynamic adjustment is:

[0227] X adjusted =λ*X real_time +(1-λ)*X historical

[0228] Among them, X adjusted is the dynamically adjusted input, and λ is the dynamic weight coefficient, which is used to balance the importance of real-time monitoring data and historical data. According to the real-time monitoring data, the value of λ can be dynamically adjusted to ensure that the model can capture the characteristics of the current state more accurately.

[0229] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.

Claims

1. A control method for a diesel generator set based on preheating time prediction, wherein the generator set includes an engine unit, a generator unit, a cooling system, a fuel system, a control system, a muffler system, an exhaust system, and a shock absorption system; the controller is used to start and stop the engine, measure, display and protect the data, and measure, display and protect the power of the generator unit; the method is characterized in that: The controller is also used to predict and control the safe preheating time of the engine unit, and then control the maximum load of the generator unit; The control method comprises the following steps: Step S101, obtaining the working information of the generator set; The working information of the generator set includes: the usage time parameters, load level parameters, maintenance status parameters, fuel quality parameters, temperature and humidity parameters and air quality parameters of the generator set; Step S102, constructing a preheating time prediction model, inputting the working information of the engine unit into the preheating time prediction model, and outputting the predicted preheating time of the engine unit; Step S103, calculating the performance degradation index of the generator set according to the above parameters, and controlling the preheating time and the maximum load of the engine unit according to the predicted preheating time of the engine unit and the performance degradation index of the generator set; Controlling the preheating time of the engine unit and the maximum load of the generator unit includes the following steps: Step S201, setting the preheating time of the engine unit according to the predicted preheating time of the engine unit output by the preheating time prediction model; Step S202, calculating the ratio of the performance degradation index of the generator set to the performance degradation index of a newly manufactured generator set; Step S203, according to the calculation result of step S202, the maximum load of the generator unit is set to a first load value, where the first load value = the maximum load of the new generator * (the ratio of the performance degradation index of the generator set to the performance degradation index of the new generator set); Methods for obtaining the working information of the generator set include: The factory time and actual recorded usage time of the generator set are used as the usage time parameter of the generator set; Obtain the maximum output power and average output power of the generator set in the latest month as load level parameters; Obtain the maintenance times of the generator set and calculate the maintenance status parameters according to the maintenance times. The calculation formula is as follows: ； The fuel purity, water content and viscosity are tested by testing equipment, and the fuel quality parameters are calculated based on the fuel purity, water content and viscosity. The calculation formula is as follows: ; In the formula, , and denote the first, second and third weight factors respectively; , and All are custom parameters; The temperature and humidity information of the area where the generator set is located is obtained as temperature and humidity parameters. The temperature and humidity information is obtained through a thermometer and a hygrometer; Obtain air quality information published by environmental monitoring stations as air quality parameters; The warm-up time prediction model includes a first hidden layer and a classifier; The working information of the engine unit is input into the first hidden layer, and the first updated feature is output; The first updated feature is input into the classifier, and the classifier outputs the predicted preheating time of the engine unit, where the predicted preheating time represents the safe preheating time of the engine unit.

2. A control method for a diesel generator set based on preheating time prediction according to claim 1, characterized in that: The first hidden layer of the warm-up time prediction model is constructed based on a multi-layer perception mechanism.

3. A control method for a diesel generator set based on preheating time prediction according to claim 1, characterized in that: Obtaining the working information of m engine units as training samples for training the preheating time prediction model; Obtain the time when the intake air temperature of each engine unit reaches the preset temperature as the sample label for training the preheating time prediction model; m is a custom parameter.

4. The control method of a diesel generator set based on preheating time prediction according to claim 1 is characterized in that: The calculation formula of the warm-up time prediction model includes: The loss function of the warm-up time prediction model is as follows: ； Where L represents the mean square error loss of the warm-up time prediction model; n represents the number of training samples; represents the predicted preheating time of the i-th engine unit output by the preheating time prediction model; represents the actual preheating time of the i-th engine unit; The gradient descent formula of the loss function of the warm-up time prediction model is as follows: ； In the formula, Represents the updated weight parameter in the warm-up time prediction model; Represents the weight parameter before updating in the warm-up time prediction model; Represents the learning rate of the warm-up time prediction model; The loss function of the warm-up time prediction model is The partial derivative of .

5. The control method of a diesel generator set based on preheating time prediction according to claim 1 is characterized in that: The calculation formula for calculating the performance degradation index of the generator set includes: ； In the formula, Indicates the performance degradation index of the generator set; represents the difference between the predicted preheating time of the engine unit output by the preheating time prediction model and the average preheating time of the engine unit newly produced; Indicates the maintenance status parameters of the generator set; Indicates the fuel quality parameters of the generator set; Indicates the air quality parameters of the generator set; , , and denote the first, second, third and fourth weight values ​​respectively; , , and These are custom parameters.

6. A control method for a diesel generator set based on preheating time prediction according to claim 1, characterized in that: In steps S201-S203, when controlling the preheating time of the engine unit and the maximum load of the generator unit, deep reinforcement learning and model predictive control intelligent algorithms are introduced to optimize the control process of the preheating time of the engine unit and the maximum load of the generator unit. The algorithm enhances accurate prediction and optimized load management, and optimizes real-time response and dynamic adjustment. The specific process is as follows: The specific process of step S201 is: Step 1: First, collect historical data of the generator set startup for training the model; these historical data include startup time, ambient temperature, humidity, and load size; represent these data as a set: in: represents the start time of the i-th generator set; Indicates the corresponding ambient temperature; Indicates ambient humidity; Indicates the load size; These data will serve as input features for training the model; Step 2: In order to simulate the actual application scenario, the collected historical data set D is divided into training data sets in chronological order And the test dataset ; Step 3: Neural network model construction, select convolutional neural network CNN as the model to solve the engine unit preheating time prediction problem; The CNN model of the network structure design includes convolutional layer, pooling layer and fully connected layer; each layer is described in detail below; Convolutional layer: The convolutional layer is used to extract features from the input data; multiple convolution kernels will be used to capture the characteristics of high-performance diesel generators at different time and space scales; convolution kernels, each of which has a size of K×K, then the output of the convolution layer can be expressed as: in, is the output of the convolutional layer, is the weight of the convolution kernel, X is the input data, is the bias term, Represents the convolution operation, and ReLU is the activation function; Known training data set obtained in step 1 The size of the input data X is × × ,in Indicates width, Indicates height, Indicates the number of channels; Convolution Kernel The size is K×K× × , where K represents the width and height of the convolution kernel, Indicates the number of convolution kernels, convolution operation X performs a convolution operation on the convolution kernel and the input data, and the resulting feature map size is × × ,in and Represent the width and height of the feature map respectively, depending on the size and step size of the convolution kernel, and then add the result of the convolution operation to the bias term Add them together and perform nonlinear transformation through the ReLU activation function to get the final convolutional layer output; Pooling layer: The pooling layer is used to reduce the dimension of the feature map while retaining the most important features; the maximum pooling operation is selected; the size of the pooling layer is F×F, then the output of the pooling layer can be expressed as: in, is the output of the pooling layer, and MaxPool represents the maximum pooling operation; Known training data set obtained in step 1 The size of the input data X is × × ,in Indicates width, Indicates height, Indicates the number of channels, The size of the pooling layer is F×F, which represents the size of each pooling window. For each pooling window, the maximum pooling operation will select the maximum value in the window as the output. The step size of the pooling operation is S, so the size of the feature map after pooling is: in, and Respectively represent the width and height of the feature map after pooling; For each pooling window, the max pooling operation selects the maximum value within the window as the output; therefore, the output of the pooling layer can be expressed as: in, Represents the value of the i-th row, j-th column, and k-th channel output by the pooling layer, represents the area corresponding to the pooling window in the input feature map, and k represents the number of channels of the input feature map; Finally, the output of the pooling layer is obtained , where the maximum value of each pooling window in the input feature map is retained; The fully connected layer is used to connect the outputs of the convolutional layer and the pooling layer and generate the final prediction result; the output of the fully connected layer is directly used as the predicted warm-up duration; the output size of the fully connected layer is M, then the output of the fully connected layer can be expressed as: in, is the output of the fully connected layer, is the weight of the fully connected layer, is the bias term, Softmax is the activation function; Output of the pooling layer The size of is W×H×C, where W represents width, H represents height, and C represents the number of channels. The output of the pooling layer is flattened into a vector of size N, where N=W×H×C; Weights of the fully connected layer The size of is N×M, where M represents the output size of the fully connected layer; The output of the fully connected layer can be expressed as the result of matrix multiplication: in, is the bias term of the fully connected layer, with dimension M; The output of the fully connected layer is processed by the activation function Softmax to convert it into a probability distribution; the definition of the Softmax function is as follows: Among them, z i represents the i-th element of the input vector, and M represents the dimension of the vector; Finally, the output of the fully connected layer is obtained , each element of which represents the probability of the corresponding category; After defining the network structure, the training data set Input into the model for training; during the training process, continuously adjust the model parameters to make the prediction results as close as possible to the actual warm-up time; finally, use the test data set to evaluate the performance of the model and check its generalization ability on new data; Step 4: Deep reinforcement learning training. The deep Q network DQN is an algorithm that combines deep learning and reinforcement learning. In this step, the DQN algorithm is selected to train the convolutional neural network CNN model, and CNN is used to approximate the Q value function to solve the problem of engine unit preheating time prediction; Design a reward function to guide the model to learn the appropriate warm-up time, so that the model can get a higher reward when the predicted warm-up time is close to the actual time, so as to gradually approach the optimal prediction through repeated training; the reward function R(s,a) is defined as the negative value of the difference between the actual warm-up time and the expected warm-up time, that is: in: is the actual preheating time; Expected warm-up time; Initialize the weights and bias items of the convolutional neural network: The weight θ of the convolution kernel and the bias item b are randomly initialized; for the initialization of the convolution layer, pooling layer and fully connected layer, the Xavier initialization method is used; Construct a DQN framework that includes a target network and an online network; the online network is used to generate Q values, and the target network is updated after a certain number of steps to provide a more stable target value. The parameters of the target network Copy from the online network parameters θ and update every fixed number of steps to prevent oscillation during training; The experience replay pool is used to store experience samples (s, a, r, s′) generated by the interaction between the model and the environment, namely the current state, action, reward, and next state; these samples will be randomly extracted during the training process for model update. Before starting deep reinforcement training, we first obtain the current state s from the environment, that is, the current start time of the generator set, ambient temperature, humidity and load size; these features constitute the state space, and then according to -greedy strategy selects action a, which is the predicted warm-up time; exploration rate Determines whether the model chooses random actions or the optimal action based on the current strategy; as training progresses, Gradually reduce; when the optimal action is selected, execute action a and observe the new state s′ and reward r; the new state s′ is the new feature of the generator after executing action a, and the reward r is calculated by the reward function R(s,a); then store the experience samples (s,a,r,s′) in the experience replay pool; these samples will be used for subsequent training; Randomly draw a mini-batch of samples from the experience replay pool ; These samples are used to calculate the target Q value and update the parameters of the online network; Calculate the target Q value: in: γ is a discount factor that weighs the impact of immediate rewards and future rewards; are the parameters of the target network, In the new state Select the Q value of the optimal action; Calculate the loss function: E is the dependent variable of the loss function. The loss function measures the gap between the target Q value and the current Q value. The parameter θ of the online network is optimized by minimizing the loss function. Update the parameters θ of the online network through the gradient descent method; calculate the gradient and update the parameters through the back propagation algorithm, and copy the parameters of the online network to the target network every fixed step. ; Q value update formula: in: α is the learning rate, which controls the step size of parameter update; γ is a discount factor that weighs the impact of immediate rewards and future rewards; θ is the parameter of the online network; are the parameters of the target network; Parameter variable explanation State s: the current start time of the generator set, ambient temperature, humidity, load size and other characteristics; these characteristics serve as inputs to the state space and constitute a description of the environment; Action a: predicted warm-up duration; each action in the action space corresponds to a predicted value of warm-up duration; Reward r: calculated by the reward function R(s,a), used to evaluate the difference between the predicted warm-up time and the actual warm-up time; Experience samples ((s,a,r,s')): including the current state, action, reward and next state. These samples are used for model training to improve the prediction performance of the model; Target Q value : In new state The Q value of the optimal action is selected to calculate the loss function and optimize the parameters of the online network; Discount factor γ: used to weigh the impact of immediate rewards and future rewards, usually between 0 and 1; Learning rate α: controls the step size of parameter updates and affects the convergence speed and stability of the model; Step 5: Predict the preheating time. Before the generator set is started, the preheating time needs to be predicted based on the current environmental conditions and historical data to ensure that the generator set reaches the best working state when it starts. By inputting these input data into the trained CNN model, the model can output the optimal preheating time. Input data: Current environmental conditions: Ambient temperature ,humidity , Current startup time , load size; historical data: historical warm-up time , Startup time , environmental conditions records; The above data are preprocessed and the input data is standardized to ensure that they have similar scales to avoid certain features having a greater impact on model training. The historical data is processed with a sliding window to extract time series features, the conditions of the past few startups, and the trends of environmental conditions; they are input into the trained CNN model, which extracts and learns features from the input data through a series of convolutional layers, pooling layers, and fully connected layers; Warm-up time for model output predictions , which indicates the optimal preheating time predicted based on the current environment and historical data, and the model prediction value , adjust the preheating time of the engine unit: if Larger: Increase the preheating time to ensure that the engine reaches the optimal operating temperature and prevent the engine from losing efficiency or damaging the equipment due to insufficient temperature during the startup process. Smaller: Reduce preheating time, save energy, and avoid unnecessary long preheating time; The specific process of step S202 is: Step 1: Collect and process the historical operating data of the generator set and extract some key features to facilitate the subsequent calculation of the performance degradation index and formulation of the optimization strategy; To facilitate the subsequent calculation of performance degradation index and formulation of optimization strategies; Usage time usage_time: For the generator set, the usage time is obtained by monitoring the start and stop time of the equipment to record the usage time of the generator set, including the total running time and the running time of each start; Load history load_history: monitors load data in real time by connecting sensors or monitoring systems, and records the load conditions of the generator unit in different time periods, including the size and changes of the load; Maintenance record maintenance_record: Information is recorded in the maintenance log or repair record, which can be collected through manual records or maintenance management systems; records the maintenance history of the generator set, including the time, content and frequency of maintenance; Running time running_time: calculate the total running time of the generator set, which can be calculated by the usage time usage_time; Load variation trend load_variation_trend: Use time series analysis method to analyze load history load_history, extract load variation trend, load fluctuation frequency and amplitude; Step 2: Build a deep Q network DQN deep reinforcement learning model and train it with historical data to obtain an accurate performance degradation index; Deep reinforcement learning model construction: The state space state takes the key features of the generator set as the state, the running time running_time and the load change trend load_variation_trend as the input of the deep Q network; the action space action: represents the load setting of the generator unit; The output is a performance degradation index, which indicates the degree of performance degradation of the generator set; Design deep neural network structures, including convolutional layers, pooling layers, and fully connected layers, as well as corresponding activation functions and loss functions; Use the historical operating data of the generator sets collected previously, including usage time, load history, and maintenance records, as training data for the model; input the historical data into the deep reinforcement learning model, and use the back-propagation algorithm and Adam optimizer to update the parameters of the model to minimize the loss function; The present invention uses the mean square error (MSE) loss function to update the parameters of the model by optimizing the loss; The specific loss function can be expressed as: in is the target Q value, calculated by the Bellman equation; According to the back-propagation algorithm and Adam optimizer, the parameters of the deep Q network are updated so that the model can better fit the training data and reward function; Define a reward function so that the model can learn the appropriate performance degradation index; the reward function should take into account the performance changes of the generator set, giving negative rewards when the performance degrades significantly, and positive rewards when the performance is stable or improved; Design the reward function R so that the model can obtain the appropriate reward value according to the current state and action; the specific reward function can be expressed as: R(state,action)=f(state,action) Performance evaluation in practical applications: Input the current generator set status into the trained deep reinforcement learning model to obtain the corresponding performance degradation index; According to the calculated performance degradation index, corresponding optimization measures are taken to adjust the load setting and maintenance frequency of the generator unit to minimize the performance degradation; this provides an accurate basis for the maximum load setting in the subsequent step S203; The dynamic adjustment process of the maximum load in step S203 is: Step 1: The model predictive control algorithm uses historical data and current status to predict the future status and performance changes of the generator unit, and dynamically adjusts the maximum load based on the prediction results to enable the generator unit to achieve the best working state while ensuring safety; Design a model predictive control algorithm and use various time series prediction methods to predict the future state and performance changes of the generator set; select the ARIMA model, whose basic formula is: in, is the observed value of the time series, c is a constant, is the autoregressive coefficient is the moving average coefficient, is the white noise error term; Combined with deep reinforcement learning algorithms, a neural network model is constructed that can learn and optimize the maximum load setting; the model will learn the optimal maximum load setting strategy based on the current generator unit state and the predicted future state to maximize the performance of the generator unit; A deep Q network DQN model is used to learn and optimize the maximum load setting; its update rule has been detailed in the previous steps, where the update formula of the Q value is as follows: Among them, Q(s,a) is the Q value of taking action a in state s, α is the learning rate, which controls the step size of parameter update, R(s,a) is the reward function, which represents the reward obtained after taking action a in state s, and γ is the discount factor, which weighs the impact of immediate rewards and future rewards. is the Q value of selecting the optimal action in the new state s′; Step 2: In the load optimization process, the recommended maximum load setting is obtained by inputting the current generator unit status into the neural network model; then, according to the safety requirements, the maximum load is dynamically adjusted to ensure that the generator unit operates in the optimal working state; Input the current state into the neural network model, which is trained by a deep reinforcement learning algorithm. By learning historical data and the current state, it can generate the optimal maximum load setting. The current generator unit state is represented as a vector S, which contains various performance indicators, temperature, pressure, and vibration; then, this state vector is used as the input of the neural network model; The neural network model processes the current state and outputs a recommended maximum load setting value ; Before applying the recommended maximum load setting, you need to ensure that it will not cause overloading of the generator unit or other safety issues; set the maximum temperature and maximum pressure As a security constraint; If the recommended maximum load setting If the safety requirements are met, it is applied to the generator unit, that is, = ; Otherwise, it is necessary to further adjust the maximum load or take other measures to ensure safety; The process of dynamically adjusting the maximum load can be expressed by the following formula: in, is the adjusted maximum load setting, is the maximum load setting that meets safety requirements; Step 3: In the performance monitoring phase, the operating status and performance indicators, temperature, pressure, and vibration of the generator set are monitored in real time, and the input of the neural network model is adjusted according to the real-time monitoring data to ensure that the model can make accurate predictions about the current status; this can achieve timely feedback and adjustments to cope with changes and emergencies in the operation of the generator set; Install sensors in the generator set to collect data on various performance indicators in real time. Temperature, pressure, and vibration sensors will continuously monitor the operating status of the generator set and send the real-time collected data to the monitoring system; The real-time monitoring data is used as one of the inputs of the neural network model so that the model can perceive the current operating status. The neural network model adopts the convolutional neural network CNN structure, and its input includes current status data and historical data. Therefore, the real-time monitoring data will constitute the input of the neural network model together with the historical data. The neural network model uses historical data and current status to predict the future status and performance changes of the generator set. In order to ensure that the model can make accurate predictions about the current status, we need to dynamically adjust the input of the neural network model according to the real-time monitoring data. Real-time monitoring data is , the historical data is , then the input of the neural network model can be expressed as X=[ , ]; The adjustment of real-time monitoring data mainly includes the following steps: Preprocess the real-time monitoring data, including data cleaning, denoising, and normalization, to improve the quality and reliability of the data; the preprocessing process can be expressed as: =Preprocess( ) The pre-processed real-time monitoring data is integrated with the historical data to form the complete input of the neural network model. The purpose of this is to use the information of the historical data to assist the model in making predictions and improve the generalization ability of the model. X=[ , ] Dynamic input adjustment: Dynamically adjust the input of the neural network model according to the changes in real-time monitoring data; if there are sudden changes or abnormalities in the real-time monitoring data, adjust its weight higher to ensure that the model can more accurately capture the characteristics of the current state; Enter the dynamic adjustment formula The dynamic weight coefficient is λ, and the input of the neural network model after dynamic adjustment is: =λ* +(1−λ)* in, is the dynamically adjusted input, and λ is the dynamic weight coefficient, which is used to balance the importance of real-time monitoring data and historical data. According to the real-time monitoring data, the value of λ can be dynamically adjusted to ensure that the model can capture the characteristics of the current state more accurately.

Images