A distributed control system and method for responding to a generator set

Abstract

This invention relates to the field of distributed control technology, solving the technical problem that the lack of a mechanism combining model prediction and neural network self-learning in existing technologies leads to a decline in control quality over operating time, severely restricting the safety and economy of generator sets. Specifically, it relates to a distributed control system and method for generator sets. The method involves: collecting the physical quantities of the target generator set; obtaining a standardized dataset of the generator set through multi-channel standardization based on these physical quantities; using a nonlinear dynamic model for rolling time-domain prediction to improve the speed of regulation and control, significantly reducing dynamic deviation; and enabling weak generator sets to maintain stable operation even when faults worsen. By adaptively adjusting the output error weighting matrix based on health status, early warning and protection are achieved, avoiding sudden damage. This method significantly reduces the rate of unplanned downtime.

Description

Technical Field

[0001] This invention relates to the field of distributed control technology, and in particular to a distributed control system and method for generator sets. Background Technology

[0002] Distributed control is a decentralized control system that connects multiple independent controllers through a communication network, enabling them to cooperate in completing complex tasks. Existing technologies for generator set control first generate a control objective function, then achieve consistent operation by obtaining operating information from adjacent units through inter-unit communication connections and coordinating to achieve dynamic balance in the output of each unit, then generate a deviation signal with a droop curve, and after the system enters steady state, calculate the deviation based on the frequency recovery to the rated value, and finally eliminate the frequency deviation. The output of each unit is adjusted by combining the control objective function and the deviation signal.

[0003] Existing technologies lack a mechanism that combines model prediction with neural network self-learning, making it impossible to optimize control parameters. This leads to a decline in control quality over time, delayed response of the protection system, and serious constraints on the safety and economy of generator sets. In practice, for example, a gas turbine may have good PID parameters when running at full load, but when the load drops to 40%, the same parameters cause a speed overshoot of up to 12%, with fluctuations lasting for 20 seconds, which seriously affects the quality of power generation. In special cases where the unit experiences slight rotor imbalance, the alarm may not be triggered due to the good performance of the PID parameters, which in turn leads to increased bearing wear and shutdown. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a distributed control system and method for generator sets. It solves the technical problems of existing technologies, which lack a mechanism that combines model prediction and neural network self-learning, making it impossible to optimize control parameters, resulting in a decline in control quality over operating time, and a lag in the response of the protection system, which seriously restricts the safety and economy of generator sets.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a distributed control method for generator sets, the method comprising the following steps: collecting the physical quantities of the target generator set, and obtaining a standardized dataset of the generator set by performing multi-channel standardization processing based on the physical quantities of the generator set; Spatiotemporal fault feature processing is performed based on the standardized generator set dataset to obtain second-depth fault features. The second-depth fault features are then evaluated using expectation-maximization confidence to generate anomaly confidence and health status assessment results. Multi-source information reward processing is performed on the standardized dataset of generator sets and the health status assessment results to obtain the second unit strategy action. Based on the health status assessment results and the second unit strategy action, the target power allocation scheme is obtained through topology consistency weighting. Based on the standardized dataset of generator sets and the target power allocation scheme, the reference control input is obtained through dynamic time-domain execution processing. Based on the reference control input, the control command set is obtained through online adaptive instruction set synthesis processing. Distributed effect feedback processing is performed based on the control instruction set to generate runtime feedback data; The process involves dynamic time-domain processing based on the standardized generator set dataset and the target power allocation scheme, including: building a dynamic model based on the standardized generator set dataset and the target power allocation scheme to obtain the predicted state trajectory; The first control sequence is obtained by performing constraint control optimization based on the predicted state trajectory, the target power allocation scheme, and the health status assessment results. The first control sequence is subjected to rolling time-domain execution processing to obtain the reference control input.

[0006] Preferably, the physical quantities of the generator set are standardized through multi-channel processing, including: The original signals from the generator set sensors are obtained by synchronously processing the physical quantities of the generator set through multiple channels. Adaptive anomaly detection and correction processing is performed on the original signals from the generator set sensors to generate generator set cleaning signals. The generator set cleaning signal is standardized to obtain a standardized generator set dataset.

[0007] Preferably, spatiotemporal fault feature processing is performed based on a standardized generator set dataset, including: The first deep fault features are obtained by extracting spatiotemporal features from the standardized generator set dataset; Based on the first-depth fault features, a preliminary fault type determination is performed, and a preliminary fault type determination result is generated. Deep fault feature processing is performed on the standardized dataset of generator sets and the preliminary fault type discrimination results to obtain the second deep fault features.

[0008] Preferably, the expected-maximization confidence assessment of the second-depth fault characteristics includes: Based on the characteristics of the second-depth fault, a Gaussian mixture model parameter set is generated through expectation maximization. Based on the parameter set of the Gaussian mixture model and the second-depth fault features, the fault type and severity are determined to obtain the fault type and fault severity. Anomaly confidence assessments are performed on fault types, Gaussian mixture model parameter sets, and second-depth fault features to generate anomaly confidence and health status assessment results.

[0009] Preferably, multi-source information reward processing is performed on the standardized generator set dataset and health status assessment results, including: The model observation status is obtained through multi-source information fusion processing based on the standardized generator set dataset and health status assessment results; Based on the model observation state, a deep-determined gradient processing is performed to obtain the first unit's strategy action; The strategy actions of the first unit, the standardized dataset of the generator unit, and the health status assessment results are processed with multi-objective compliance rewards to obtain the strategy actions of the second unit.

[0010] Preferably, the process involves topological consistency weighting based on health status assessment results and second unit strategy actions, including: Topological consistency processing is performed based on the standardized generator set dataset and health status assessment results to generate consistent state variables; Distributed Newton optimization is performed based on consistent state variables to obtain the Newton optimization allocation solution. An adaptive weighted fusion of the second unit's strategy actions and the Newton-optimal allocation solution is performed to obtain the target power allocation scheme.

[0011] Preferably, online adaptive instruction set synthesis processing based on reference control input includes: Adaptive PID parameters are obtained through online adaptive processing based on the reference control input, health status assessment results, and generator set standardized dataset; The second PID parameters are obtained by performing back error gradient descent processing based on adaptive PID parameters, target power allocation scheme and generator set standardized dataset; The second PID parameters, the target power allocation scheme, and the health status assessment results are processed by instruction set integration to obtain the control instruction set.

[0012] Preferably, distributed effect feedback processing based on the control instruction set includes: Based on the control instruction set, distributed instruction processing is used to obtain the generator set node control instructions; Based on the generator set node control commands, the generator set control execution process is performed to obtain the generator set execution status; The execution status of the generator set is processed to provide feedback on the execution effect, and operational feedback data is generated.

[0013] Preferably, the physical quantities of the generator set include generator set vibration acceleration, generator set temperature, generator set current, generator set voltage, generator set frequency, generator set active power, generator set reactive power, and generator set speed.

[0014] This technical solution also provides a distributed control system for generator sets, which includes: The standardization module is used to collect the physical quantities of the target generator set and obtain a standardized dataset of the generator set through multi-channel standardization processing based on the physical quantities of the generator set. The health assessment module is used to perform spatiotemporal fault feature processing based on the generator set standardized dataset to obtain second-depth fault features, perform expectation-maximization confidence assessment on the second-depth fault features, and generate abnormal confidence and health status assessment results. The power allocation module is used to perform multi-source information reward processing on the standardized dataset of generator sets and the health status assessment results to obtain the second unit strategy action. Based on the health status assessment results and the second unit strategy action, the target power allocation scheme is obtained through topological consistency weighting. The control command module is used to obtain reference control input by dynamically executing the time domain based on the generator set standardized dataset and target power allocation scheme, and to perform online adaptive command set synthesis processing based on the reference control input to obtain the control command set; The runtime feedback module is used to perform distributed effect feedback processing based on the control instruction set and generate runtime feedback data.

[0015] By employing the above technical solution, the present invention provides a distributed control system and method for generator sets, which has at least the following beneficial effects: 1. The fault diagnosis and anomaly identification method constructed in this invention not only achieves high-precision identification of generator set fault types, but also avoids the problems of blind manual parameter tuning and poor adaptability to operating conditions through adaptive optimization of hyperparameters, significantly improving the robustness of the model under different operating conditions. By extracting spatiotemporal joint features and combining them with high anomaly confidence corresponding to low-density regions, it effectively realizes the detection of unknown faults, greatly reducing the false negative rate. The joint output of anomaly confidence and fault severity extends the lifespan of the entire generator set cluster.

[0016] 2. This invention avoids the single-point failure risk of centralized scheduling by using multi-agent deep reinforcement learning, significantly reducing power tracking error and the total cost of power generation. Through distributed Newton optimization, under the premise of meeting the total load demand, the severity of the fault is used as a weighting factor in the cost function to provide a conservative global optimal solution. The fusion of the two achieves a balance between aggressive exploration and conservative safety, retaining the efficiency of deep reinforcement learning while avoiding the risk of weak units being allocated too much power. Furthermore, by making the load allocation priority entirely determined by the health status index, healthy units take on more load, which greatly reduces the rate of deterioration of the unit's health status.

[0017] 3. This invention uses a nonlinear dynamic model for rolling time-domain prediction, anticipating load change trends and optimizing control sequences in advance. This solves the lag problem of large inertial systems, improves regulation and control speed, and significantly reduces dynamic deviation. It adaptively adjusts the output error weighting matrix based on health status, while a BP neural network adjusts PID parameters online based on health status assessment results. This synergy allows weak units to maintain stable operation even when faults worsen. By adaptively adjusting the output error weighting matrix based on health status and dynamically reducing protection action thresholds based on health status, early warning and protection are achieved, preventing sudden damage. In simulation tests, this method significantly reduces the unplanned downtime rate.

[0018] 4. This invention uses point-to-point or broadcast methods to send instruction sets to each unit, avoiding the risk of all units going out of control due to a single point of communication failure in centralized dispatching. At the same time, each local node only extracts its own instructions, which greatly reduces the communication bandwidth usage. It also uses amplitude limiting processing to ensure that the actuators operate within a safe range. The actual operating parameters of the units are collected in real time through a closed-loop feedback monitoring method, and these data are used as the input for the next control cycle S1. The closed-loop feedback mechanism enables the system to cope with interference such as sudden changes in grid load and sensor noise, greatly reducing the frequency fluctuation range and significantly improving the power generation quality and operational stability. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a distributed control method for generator sets according to the present invention; Figure 2 This is a structural block diagram of a distributed control system for generator sets according to the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] Because existing technologies lack a mechanism that combines model prediction with neural network self-learning, they cannot optimize control parameters, leading to a decline in control quality over operating time and a lag in the response of the protection system. This severely restricts the safety and economy of generator sets. Please refer to [the relevant documentation / reference]. Figure 1 This embodiment provides a distributed control method for generator sets, which can dynamically optimize control parameters, ensure stable control quality, prevent system response lag, and improve the safety and economy of generator sets. The method includes the following steps: S1. Collect the physical quantities of the target generator set, and obtain the standardized dataset of the generator set through multi-channel standardization processing based on the physical quantities of the generator set. The existing technology lacks a signal processing model that organically links the three links of acquisition, cleaning and standardization. Each link operates independently and the parameters are fixed. It is impossible to achieve data consistency across channels, operating conditions and generator sets while ensuring data authenticity. This seriously restricts the reliability and adaptability of fault diagnosis, load distribution and parameter tuning in the distributed control system. To solve the above problems, the specific implementation steps are as follows: S11. Based on the physical quantities of the generator set, the original signals of the generator set sensors are obtained through multi-channel synchronous processing. In this step, for each sampling channel of each generator set, the sensor gain coefficient corresponding to the channel is multiplied by the physical quantity of the generator set at the current sampling time of the channel to obtain a product value. Then, the obtained product value is added to the sensor bias of the channel. After the above multiplication and addition operations, the original signal of the generator set sensor at the sampling channel and the sampling time of the generator set is finally obtained. This signal is the original electrical characterization of the actual operating physical quantities of each generator set, such as vibration, temperature, current, voltage, etc., and serves as the basic input for subsequent data cleaning and feature standardization. Through this acquisition method, the original signals of the generator set sensors of all generator sets can be acquired synchronously. Among them, the physical quantities of the generator set include, but are not limited to: generator set vibration acceleration, generator set temperature, generator set current, generator set voltage, generator set frequency, generator set active power, generator set reactive power, and generator set speed.

[0022] S12. Based on the original signals of the generator set sensors, perform adaptive anomaly detection and correction processing to generate generator set cleaning signals. In this step, for each sampling channel of each generator set, take the current sampling time as the end point, take a fixed number of past sampling points to form a sliding window, add the original signals of the generator set sensors at all sampling times in the sliding window, and then divide by the width of the sliding window to calculate the mean of the original signals in the sliding window. Then, for each sampling time in the sliding window, calculate the difference between its original signal and the mean, square each difference, add all the squared values ​​and divide by the width of the sliding window. Then, the square root of the result is taken to obtain the standard deviation of the original signal within the sliding window. Next, the absolute value of the difference between the original signal at the current sampling time and the mean is calculated, and the absolute value is compared with three times the standard deviation. If the absolute value is less than or equal to three times the standard deviation, the current original signal is determined to be normal, and the current original signal of the generator set sensor is directly used as the output signal after generator set cleaning. If the absolute value is greater than three times the standard deviation, the current original signal is determined to be an outlier, and the mean within the sliding window calculated earlier is used as the output signal after generator set cleaning.

[0023] S13. Standardize the generator set cleaning signal to obtain a standardized generator set dataset. In this step, for each sampling channel of each generator set, retrieve the pre-stored historical baseline mean and historical baseline standard deviation of that sampling channel from the local intelligent control node. The historical baseline mean and historical baseline standard deviation are derived from long-term operating data statistics under the historical normal operating conditions of the generator set. Then, subtract the corresponding historical baseline mean from the generator set cleaning signal of the current sampling channel to obtain a difference. Next, divide the difference by the corresponding historical baseline standard deviation. After the above subtraction and division operations, the standardized operating dataset component of the sampling channel of the generator set is finally obtained. Through the above calculations, each sampling channel of each generator set can obtain a dimensionless standardized value at each time. These values ​​together constitute the standardized operating dataset of the generator set.

[0024] This invention ensures the temporal consistency of all sensor signals by cascading three sub-steps: signal acquisition, data cleaning, and feature standardization. While eliminating outlier noise, it retains the edge features of real fault signals, ensuring that the standardized data does not distort historical baseline statistics due to outlier interference. It also eliminates differences in the dimensions and magnitudes of different sensors, significantly improving the sensitivity of fault diagnosis and the efficiency of data preprocessing. The standardized dataset is consistent and comparable among all local intelligent control nodes, directly enhancing the accuracy and convergence speed of adaptive tuning in subsequent steps.

[0025] S2. Spatiotemporal fault feature processing is performed based on the standardized generator set dataset to obtain second-depth fault features. The second-depth fault features are evaluated with expectation-maximization confidence to generate anomaly confidence and health status assessment results. Existing technologies have difficulty in achieving both high-precision classification of known faults and reliable detection of unknown faults. The models have poor adaptability and cannot provide continuous quantitative indicators of health status, which seriously restricts the effectiveness of differentiated load allocation and control parameter self-tuning in distributed control systems. To solve the above problems, the specific implementation steps are as follows: S21. Based on the standardized generator set dataset, the first-depth fault features are obtained through spatiotemporal feature extraction. In this step, the forward propagation method of the long short-term memory convolutional neural network is adopted, which is divided into two stages: convolutional layer feature extraction and long short-term memory layer time series modeling. In the first stage, for the standardized operation dataset of each generator set, which is organized in the form of a time series matrix, for each convolutional kernel, the weight parameters of the convolutional kernel in all sampling channels and all convolutional depths are multiplied by the corresponding data values ​​to obtain multiple product values. Then, all product values ​​are added together, and the bias parameters of the convolutional kernel are added to obtain a cumulative sum. Finally, the cumulative sum is input into the activation function of the convolutional layer to calculate the output feature value of the convolutional kernel at the current time. The above process is repeated for all convolutional kernels to obtain the feature vector at the current time. In the second stage, the feature sequence output from the convolutional layer is input into the Long Short-Term Memory (LSTM) layer and calculated recursively step by step. For each time step, the forget gate is calculated first. The output of the hidden layer at the previous time step is concatenated with the current feature vector, multiplied by the forget gate weight matrix, and then the forget gate bias is added. The forget gate output is obtained by passing the sigmoid function. Similarly, the values ​​of the input gate, output gate, and candidate memory units are calculated. Then, the state of the memory unit at the previous time step is multiplied element-wise by the forget gate output, and the candidate memory unit is multiplied element-wise by the input gate output. The two are added together to obtain the state of the memory unit at the current time step. Finally, the state of the memory unit at the current time step is processed by the hyperbolic tangent function and multiplied element-wise by the output gate output to obtain the output of the hidden layer at the current time step.

[0026] S22. Based on the first deep fault features, perform preliminary fault type discrimination and generate preliminary fault type discrimination results. In this step, for each fault category, firstly, perform a dot product operation on the full connected layer weight vector corresponding to the category and the generator set deep fault feature vector, that is, multiply the corresponding positions and accumulate them to obtain a product sum. Then, add the product sum to the bias value of the category to obtain the unnormalized score of the category. Then, perform an exponential operation on the unnormalized score of each fault category to obtain the exponential value corresponding to the category. Finally, add the exponential values ​​of all fault categories to obtain a total exponential sum. Subsequently, for each fault category, its index value is divided by the sum of the total indices to obtain the probability value of the occurrence of the fault category given the current depth fault characteristics. Finally, the probability values ​​of all fault categories are compared, and the fault category corresponding to the highest probability value is taken as the preliminary fault type discrimination result of the generator set. This result is a discrete category label, which indicates the fault type initially judged based on the depth features, such as abnormal vibration, excessive temperature, etc. It can map the high-dimensional deep fault features to the fault category space to provide an analytical basis for subsequent steps and provide a loss calculation benchmark for the subsequent improved particle swarm optimization.

[0027] S23. Perform deep fault feature processing on the standardized dataset of generator sets and the preliminary fault type discrimination results to obtain the second deep fault features. In this step, the initial values ​​of the hyperparameters to be optimized, such as the learning rate, the number of convolutional kernels, and the dimension of the long short-term memory layer hidden layer, are first encoded as particle position vectors. Each particle represents a set of candidate hyperparameters. For each particle, its current velocity vector is first multiplied by the inertia weight under the current iteration number to obtain a weighted velocity value. Then, the difference between the particle's historical best position vector and its current position vector is calculated. This difference is multiplied by the individual learning factor and then by a random number to obtain the individual guidance increment. Next, the difference between the global best position vector and the current position vector is calculated. This difference is multiplied by the social learning factor and then by another random number to obtain the social guidance increment. The weighted velocity value, the individual guidance increment, and the social guidance increment are added together to obtain the particle's new velocity vector. Then, the particle's current position vector is added to the new velocity vector to obtain the particle's new position vector. In the above process of updating velocity and position, the inertia weight adopts a non-linear decreasing strategy. First, the ratio of the current iteration number to the maximum iteration number is calculated. The nonlinear adjustment factor of the ratio is raised to the power of the product, multiplied by the difference between the maximum and minimum inertia weights, and then the product is subtracted from the maximum inertia weight to obtain the inertia weight for the current iteration. After all particles are iterated and updated to the maximum number of iterations, the globally optimal position vector is taken as the optimal hyperparameter combination. Based on this, the long short-term memory convolutional neural network is reconfigured, and the forward propagation calculations of S21 and S22 are re-executed to finally obtain the optimized generator set deep fault features. The individual learning factor and the social learning factor are generally set to 2.0 and 2.0, respectively. In some improved algorithms, values ​​between 1.5 and 2.5 can also be used to balance the local search and global search capabilities. The maximum number of iterations is usually set to 100. If the model complexity is high or the dataset is large, it can be increased to 200 or 300 to ensure sufficient convergence.

[0028] S24. Based on the characteristics of the second-depth fault, a Gaussian mixture model parameter set is generated through expectation maximization. In this step, the number of Gaussian components is first set, and the mixture weight, mean vector, and covariance matrix of each component are initialized. Then, the expectation step is entered: For the second-depth fault feature vector of each generator set and each Gaussian component, the probability density value of the feature vector under the component is calculated. That is, based on the mean vector and covariance matrix of the component, the squared Mahalanobis distance of the feature vector from the mean is calculated, and then the density value is obtained by exponential operation and normalization coefficient. The density value is multiplied by the current mixture weight of the component to obtain a weighted density value. Then, the weighted density values ​​of all components are added to obtain the total density value. Finally, the weighted density value of the component is divided by the total density value to obtain the posterior probability of the feature vector belonging to the component. After calculating the posterior probabilities of all samples, the process proceeds to the maximization step. For each Gaussian component, the posterior probabilities of all generator sets are summed to obtain a probability sum. This probability sum is then divided by the total number of generator sets to obtain the updated mixture weights. Next, for each generator set, its second deep fault feature vector is multiplied by the posterior probability. All products are summed and divided by the probability sum to obtain the updated mean vector. Then, for each generator set, the difference between its second deep fault feature vector and the updated mean vector is calculated. This difference is multiplied by its own cross product (i.e., the difference multiplied by its transpose) and then multiplied by the posterior probability. All results are summed and divided by the probability sum to obtain the updated covariance matrix. The expectation and maximization steps are repeated until the parameters converge. The final output Gaussian mixture model parameter set consists of the mean vector, covariance matrix, and mixture weights of each Gaussian component.

[0029] S25. Based on the Gaussian mixture model parameter set and the second-depth fault features, determine the fault type and severity to obtain the fault type and severity. In this step, the maximum a posteriori probability classification and Mahalanobis distance quantization method are used for processing. First, the fault type is determined. For each Gaussian component, according to the component's mixing weight, mean vector, and covariance matrix in the Gaussian mixture model parameter set, the posterior probability of the second-depth fault feature belonging to that component is calculated. Specifically, for each component, the second-depth fault feature vector is subtracted from the component's mean vector to obtain the difference vector. The quadratic form of the difference vector and the inverse of the component's covariance matrix is ​​calculated, and then exponential operation is performed and multiplied by the mixing weight to obtain the weighted probability density of that component. Finally, the weighted probability densities of all components are added together to obtain the total density. Finally, the weighted probability density of each component is divided by the total density to obtain the posterior probability of that component. The posterior probabilities of all Gaussian components are compared, and the fault category corresponding to the component with the largest posterior probability is taken as the fault type of the generator set. Then, the severity of the fault is quantified. After determining the Gaussian component, the mean vector and the inverse of the covariance matrix of that component are extracted from the parameter set of the Gaussian mixture model. The mean vector of that component is subtracted from the second-depth fault feature vector to obtain the difference vector. The product of the difference vector and the inverse of the covariance matrix is ​​calculated to obtain an intermediate vector. Then, the dot product of the intermediate vector and the difference vector is performed to obtain a scalar value. Finally, the square root of the scalar value is performed to obtain the Mahalanobis distance, which is the severity of the generator set fault.

[0030] S26. Anomaly confidence assessment is performed on the fault type, Gaussian mixture model parameter set, and second-depth fault features to generate anomaly confidence and health status assessment results. In this step, a probability density threshold discrimination method is used. First, for each generator set, the overall probability density value of its second-depth fault features under the entire Gaussian mixture model is calculated. For each Gaussian component, the difference between the feature vector and the mean vector of the component is calculated. The quadratic form of the difference and the inverse of the covariance matrix is ​​calculated. After exponential operation, it is multiplied by the mixture weight of the component to obtain the weighted probability density of the component. Then, the weighted probability densities of all components are added together to obtain the overall probability density value. At the same time, the maximum probability density value in the entire feature space and the pre-set normal state baseline probability density threshold are obtained from the Gaussian mixture model parameter set. The threshold is the minimum probability density value of historical normal operation data under the Gaussian mixture model. Next, the generator set anomaly confidence level is calculated by dividing the overall probability density value by the maximum probability density value to obtain a ratio. Then, the ratio is subtracted from the value to obtain the anomaly confidence level, which is between zero and one. Then, anomaly judgment is performed by comparing the calculated anomaly confidence level with the normal state baseline probability density threshold. If the anomaly confidence level is greater than the threshold, the generator set is judged to be abnormal; otherwise, it is judged to be in normal condition. Finally, the generator set fault type and generator set fault severity obtained in sub-step 2.5 are combined with the generator set anomaly confidence level calculated in this sub-step to form the generator set health status assessment result.

[0031] This invention, through its fault diagnosis and anomaly identification method, not only achieves high-precision identification of generator set fault types, but also avoids the blindness of manual parameter tuning and poor adaptability to operating conditions through adaptive optimization of hyperparameters. This significantly improves the robustness of the model under different operating conditions. More importantly, after the optimized deep fault features are input into the GMM, the clustering boundaries of each Gaussian component of the GMM become clearer, and the severity of the fault quantified by Mahalanobis distance becomes more accurate. At the same time, the anomaly confidence of the probability density threshold discrimination is also more reliable. In practice, vibration sensors collect weak periodic pulses of early bearing wear, but normal fluctuations in the temperature channel are misjudged by the model as associated anomalies, leading to missed detections. If the generator set suddenly drops from full load to half load, the CNN with a fixed number of convolutional kernels cannot capture the feature changes of the dynamic response, and the fault identification accuracy drops significantly.

[0032] This invention extracts spatiotemporal joint features, and IPSO automatically adjusts the network structure to enhance sensitivity to slowly changing features. After the optimized features are input into GMM, the Mahalanobis distance that deviates from the normal distribution is accurately amplified, thereby accurately identifying composite fault types and quantifying their severity. By calculating the overall probability density through GMM, low-density areas correspond to high anomaly confidence, effectively realizing the detection of unknown faults and significantly reducing the false negative rate. The joint output of anomaly confidence and fault severity extends the lifespan of the entire unit cluster. In practice, special situations may arise, such as asymmetrical temperature distribution caused by coolant leakage in the generator set. This fault mode has never appeared in historical data, and traditional supervised models misjudge it as normal, while the actual fault continues to develop and eventually leads to shutdown.

[0033] S3. Perform multi-source information reward processing on the standardized dataset of generator sets and the health status assessment results to obtain the second unit strategy action. Based on the health status assessment results and the second unit strategy action, the target power allocation scheme is obtained through topology consistency weighting. Existing technologies lack a distributed, health-adaptive, and multi-objective collaborative load allocation framework, making it difficult to achieve refined protection of differentiated health units while ensuring system economy and stability. To solve the above problems, the specific steps are as follows: S31. Based on the generator set standardized dataset and health status assessment results, the model observation state is obtained through multi-source information fusion processing. In this step, for each generator set model, its own generator set standardized operation dataset component and its own health status assessment results are directly used as the first and second parts of the observation state. Then, the information of neighboring generator sets is processed. All neighboring generator sets of the generator set are obtained through the communication network. For each neighbor, the generator set standardized operation dataset component of the neighbor is multiplied by the communication weight coefficient corresponding to the neighbor to obtain a weighted dataset. Then, the weighted datasets of all neighbors are summed to obtain a neighbor weighted sum dataset. Similarly, the generator health status assessment result of each neighbor is multiplied by the corresponding communication weight coefficient to obtain the weighted health status result. Then, the weighted health status results of all neighbors are summed to obtain the neighbor weighted sum health status result. Finally, the four parts—the self-standardized dataset, the self-health status assessment result, the neighbor weighted sum dataset, and the neighbor weighted sum health status result—are concatenated in sequence to form the complete model observation state of the generator model. This observation state vector contains the real-time operating information and health status of each generator, as well as the aggregated information of neighbor generators obtained through distributed communication, enabling each model to perceive the overall situation within a local range when making decisions.

[0034] S32, Model-based observation state GSAO S Perform depth-deterministic gradient processing to obtain the first unit's strategy action (GSP). A This step employs a multi-model deep deterministic policy gradient method. Each generator unit corresponds to one model, and each model maintains a policy network and an evaluation network. First, the current generator unit model observation state (GSAO) is processed. S The input is a policy network, which internally performs linear transformations and nonlinear activation function operations through multiple fully connected layers, outputting a preliminary action vector DZ. X The action vector DZ X Includes recommended active power P Y And suggested reactive power P WG Two components; then an exploration noise vector TSZ is generated. S Typically generated using the Ornstein-Uhlenbeck process, it exhibits time correlation. The exploration noise vector TSZ is then used. S Add to motion vector DZ X Above, the first unit's strategy action GSP is obtained. A This action represents the power allocation value recommended for the generator set at the current moment. During model training, the evaluation network receives observations of the GSAO status of all generator sets. S and motion vector DZ X Output a value score JZ F This is used to guide the updating of the policy network; Output a value score JZ F The specific steps are as follows: first, calculate the first unit's strategy action (GSP). A With the weight matrix QZ of this layer A The product of these terms yields multiple product values ​​CJ. Z Then sum all the product values ​​and add the bias value PZ of that layer. aTo obtain an intermediate sum ZJ h ; this intermediate and ZJ h The input activation function, typically a linear rectified function or a hyperbolic tangent function, yields the output vector XL of the layer. B The calculation proceeds layer by layer until the output layer, which has only one node. The output value is the value score JZ given by the evaluation network. F , representing the cumulative reward expected to be obtained by taking the current joint action in the current state; during training, the parameter update direction of the policy network is to make the value score JZ output by the evaluation network more valuable. F Maximizing this is specifically achieved by calculating the gradient (DZT) of the evaluation network output with respect to the action. D Then multiply by the gradient ZSC of the policy network output with respect to its own parameters. S To achieve this, the parameter updates of the evaluation network are accomplished by minimizing the temporal difference error, i.e., by calculating the current output DQP of the evaluation network. J With target value MBJ Z The mean square error between them, where the target value MBJ Z ZKY, consisting of instant rewards plus a discount factor Z Multiply by the output MBP of the target evaluation network J get.

[0035] S33, GSP strategy action for the first unit A Generator Set Standardized Dataset GSO D and health status assessment results GSH R Perform multi-objective reward processing to obtain the second crew strategy action (DESP). A In this step, four reward components are first calculated for each generator set, including the power point tracking reward component PGZ. L During the calculation, first obtain the actual output power P at the current moment. actual and the first unit's strategy action GSP A The recommended active power P Y Subtract the two and take the absolute value, then divide by the rated power P. rated The power tracking reward component PGZ is obtained. L Negative values; economic reward component JJX L During calculation, the actual output power P is used. actual Substituting the current cost into the generation cost function Ci, and then dividing by the maximum generation cost C, yields the current cost. max Then take the negative value; Health status reward component JKF L During the calculation, the health status assessment results (GSH) are used. R Extracting fault severity GSF S and outlier confidence level GSA CFirst, the abnormal confidence level GSA C Add the value 1, then multiply by the severity of the fault, GSF. S Finally, take the negative value; stability bonus component WDJ L During the calculation, the power change ΔP between adjacent time points is calculated. i Divide by the rated power P rated Square the result and then take the negative value; then multiply the above four components by their respective weight coefficients ω1, ω2, ω3, and ω4, and finally add the four products together to obtain the instantaneous reward value r at the current moment. i The weighting coefficients can be obtained through the analytic hierarchy process (AHP). Typical practical values ​​are ω1=0.4, ω2=0.2, ω3=0.3, and ω4=0.1. Finally, the instantaneous reward value r... i Substituting these values ​​into the update formulas for the policy network and evaluation network parameters in S32, including minimizing the current evaluation network loss function and calculating the deterministic policy gradient of the policy network, the network parameters are iteratively updated through backpropagation and gradient descent until convergence, yielding the optimized second unit policy action (DESP). A .

[0036] First, based on the input of the first unit policy action GSP A Standardized dataset GSO D and health status assessment results GSH R The system calculates four reward components: power tracking error, economic cost, health status cost, and power fluctuation penalty. Each component is negative, representing a penalty term. These four components are then multiplied by preset weight coefficients and summed to obtain the instantaneous reward value. Finally, this instantaneous reward value is substituted into the policy network and evaluation network update formulas in S32. This includes minimizing the evaluation network loss and calculating the policy network gradient. The network parameters are iteratively updated through backpropagation and gradient descent to optimize the policy network's output actions towards higher rewards, ultimately outputting the second unit's policy action, DESP. A The reward function uses a weighted method to linearly weight and fuse the four components into a single scalar reward value, which is the most direct and mature processing method in multi-objective reinforcement learning. For the acquisition of weight coefficients, the analytic hierarchy process (AHP) is a recognized practice in academia. In the reward value calculation of Q-learning, AHP is explicitly used to determine the weight of each objective. Combining AHP with reinforcement learning is used for equipment health status evaluation. In the design of each component, power tracking, economy, health status, and stability all have clear academic sources: the design methods such as the absolute value normalization of power tracking deviation, the normalization of economic cost, the fusion of fault severity and confidence in health status, and the squared penalty term for power change can all be verified and supported by relevant research in this field.

[0037] S34. Perform topology consistency processing based on the generator set standardized dataset and health status assessment results to generate consistent state variables. First, for each generator set, determine the available adjustable capacity based on the current adjustable capacity information in its standardized operating dataset. At the same time, extract the fault severity and anomaly confidence from the health status assessment results. First, calculate the value of adding the anomaly confidence, then multiply the value by the fault severity to obtain an intermediate product. Then, divide the value by the intermediate product to obtain the health status index. Combine the available adjustable capacity and the health status index to form the initial consistent state variables of the generator set. Then, each generator set obtains the current consistent state variables of all neighboring generator sets through the communication network and performs consistent iterative updates. For each neighbor, first calculate the difference between the neighbor's state variable and its own state variable, then multiply the difference by the communication connection weight corresponding to the neighbor to obtain a weighted difference; sum the weighted differences of all neighbors to obtain the total weighted difference; multiply the total weighted difference by the consistency step size factor to obtain the update increment; finally, add the update increment to its own current state variable to obtain the updated consistent state variable. Repeat the above iterative process until the available regulation capacity and health status index of each generator set reach consistency within the communication topology, and finally output the generator set consistent state variables of each generator set.

[0038] S35. Perform distributed Newton optimization based on consistent state variables to obtain the Newton optimization allocation solution. In this step, for each generator set, the health weight factor is calculated based on the health state index in its consistent state variables. Specifically, the fault severity and anomaly confidence are extracted from the health state assessment results. First, the value of 1 plus the anomaly confidence is calculated, then multiplied by the fault severity, and then 1 is added again to obtain the health weight factor of the generator set. Then, the optimization objective is to minimize the weighted sum of the power generation cost of each generator set and the health weight factor, while satisfying the total load demand equal to the sum of the power of each generator set and the upper and lower limits of the power of each generator set. The solution process is as follows: Initialize the power allocation value and the local estimate of the Lagrange multiplier for each generator set to zero, and enter the iteration loop. For each generator set, first calculate the first derivative and the second derivative of the power generation cost function under the current power; then calculate the first derivative and subtract the current local estimate of the Lagrange multiplier to obtain the difference. The second derivative is then calculated and a regularization parameter is added to obtain the denominator. The difference is divided by the denominator to obtain a correction amount. The current power value is subtracted from the correction amount to obtain the preliminary new power value. This preliminary new power value is compared with the lower and upper power limits. If it is less than the lower limit, the lower limit is taken; if it is greater than the upper limit, the upper limit is taken; otherwise, it remains unchanged to obtain the updated power allocation value. At the same time, for the update of the Lagrange multipliers: first, the self-multiplier is calculated, plus the consistency step size multiplied by the weighted sum of the differences of all neighboring multipliers, plus the power imbalance correction step size multiplied by the difference between the sum of the power of all units and the total load demand to obtain the updated multiplier estimate. The above iterative update of power and multipliers is repeated until the power change of all units is less than the preset threshold and convergence occurs. The converged power allocation value of each generator unit is the generator unit Newton optimization allocation solution. This solution achieves economic allocation considering the healthy state while meeting the total load demand.

[0039] S36. Adaptively weighted fusion of the second unit's strategy action and Newton's optimization allocation solution is performed to obtain the target power allocation scheme. In this step, for each generator unit, the adaptive fusion weight coefficient is first calculated based on its health status assessment results. Specifically, the fault severity and anomaly confidence are extracted from the health status assessment results. The value of the anomaly confidence is calculated and then multiplied by the fault severity to obtain a product value. This product value is multiplied by the attenuation factor and then negative. The product value is then exponentially calculated with the natural constant as the base to obtain the exponent value. The difference between the maximum weight coefficient and the minimum weight coefficient is calculated. This difference is multiplied by the exponent value and then added to the minimum weight coefficient to obtain the adaptive fusion weight coefficient of the generator unit. The characteristic of this coefficient is that the worse the health status, the smaller the coefficient. Then, the suggested active power in the second unit's strategy action is multiplied by the aforementioned weighting coefficient to obtain the weighted strategy power; the Newton optimization allocation solution is multiplied by one minus the weighting coefficient to obtain the weighted Newton power; the two are added together to obtain the target active power of the generator unit. Next, the load allocation priority of each generator unit is calculated: first, the health status index of each generator unit is calculated based on the health status assessment results, then one is divided by one plus the fault severity multiplied by one plus the anomaly confidence, and then the health status index of each unit is divided by the sum of the health status indices of all units to obtain the load allocation priority of that unit. Finally, the target active power and target reactive power of each generator unit are calculated using a dual-model fusion method similar to that for active power, and the load allocation priority is combined to form a complete generator unit target power allocation scheme.

[0040] This invention utilizes multi-agent deep reinforcement learning to enable each generator unit to make autonomous decisions based on its own and its neighbors' real-time status. This avoids the single-point failure risk of centralized dispatching, significantly reduces power tracking error, and substantially lowers the total cost of power generation. In practice, for example, a power plant uses a central controller to uniformly allocate loads. When the network interface of this controller fails, the entire system collapses, causing 10 generator units to be unable to respond to grid dispatch commands. This results in the frequency dropping from 50Hz to 49.2Hz, triggering low-frequency load shedding and the shedding of loads.

[0041] This invention utilizes distributed Newton optimization to incorporate a health status index, varying according to the severity of the fault, into the cost function as a weighting factor, while meeting total load requirements. For example, when a unit's fault severity is 0.6 and its anomaly confidence is 0.8, its health weighting factor increases from 1 to 1 + 0.6 × 1.8 = 2.08, causing the unit's allocated power to automatically decrease by approximately 30%, thus preventing overloading of faulty units. This effect complements multi-agent deep reinforcement learning, which aims to maximize long-term rewards but may overuse healthy units in the short term. Distributed Newton optimization, on the other hand, provides a conservative global optimal solution. The fusion of the two achieves a balance between aggressive exploration and conservative safety.

[0042] This invention combines distributed Newtonian optimization with multi-agent deep reinforcement learning, retaining the high efficiency of deep reinforcement learning while avoiding the risk of weak units being allocated excessive power. Furthermore, by determining load allocation priority entirely by the health status index, healthy units take on more load, while weak units automatically reduce their load. In long-term operation tests, such as simulating 2000 hours, this method significantly reduces the rate of deterioration in the health status of units, effectively reducing the number of unplanned outages. In real-world cases, existing technologies focus solely on minimizing power generation costs, ignoring health costs. This leads to a low-cost, older unit operating at full load for extended periods, resulting in a threefold increase in failure rate within two years, with maintenance costs far exceeding the fuel savings.

[0043] S4. Based on the standardized dataset of the generator set and the target power allocation scheme, the reference control input is obtained through dynamic time-domain execution processing. Based on the reference control input, the control command set is obtained through online adaptive instruction set synthesis processing. The existing technology lacks a mechanism that combines model prediction and neural network self-learning, and cannot optimize control parameters in real time according to the health status and operating conditions of the unit. This leads to a decline in control quality over time, a lag in the response of the protection system, and seriously restricts the safety and economy of the generator set. To solve the above problems, the specific steps are as follows: S41. Based on the standardized dataset of generator sets and the target power allocation scheme, a dynamic model is constructed to obtain the predicted state trajectory. In this step, a discrete-time nonlinear state-space model is first established for each generator set. This model describes the state vector, including boiler steam pressure, turbine speed, generator output power, etc. How to calculate the state vector, control input vector, fuel valve opening, steam valve opening, excitation voltage, etc., and measurable disturbance vector, such as changes in grid load demand, at the next moment through the state transition function. At the same time, the output vector is calculated from the current state vector through the output function. Based on the above model, rolling predictions are made for multiple future moments: taking the actual state estimate at the current moment as the starting point, for each prediction step in the prediction time domain, the control input sequence and measurable disturbance sequence from the current moment to the moment before the prediction step are obtained first. Then, starting from the current state, the state transition function is applied sequentially: First, the current state, the first-step control input, and the first-step disturbance are substituted into the state transition function to obtain the first-step predicted state; second, the first-step predicted state, the second-step control input, and the second-step disturbance are substituted into the state transition function to obtain the second-step predicted state; and so on, recursively calculating until the last moment of the prediction time domain is reached. Each application of the state transition function actually involves multiplying and adding multiple variables, such as multiplying the fuel valve opening by the gain coefficient and adding it to other terms like the steam valve opening, as well as combining nonlinear functions. The final result, all the predicted states from the first step to the second prediction step, arranged in chronological order, constitute the generator set's predicted state trajectory.

[0044] S42. Based on the predicted state trajectory, target power allocation scheme, and health status assessment results, constraint control optimization is performed to obtain the first control sequence. In this step, the output error weighting matrix is ​​dynamically adjusted according to the health status assessment results: the fault severity and anomaly confidence are extracted, and the value of adding the anomaly confidence is calculated first, and then multiplied by the fault severity to obtain the product value; the product value is multiplied by the health status influence coefficient to obtain the adjustment amount; then the baseline output error weighting matrix is ​​multiplied by the adjustment amount to obtain the adaptive weighting matrix, and then the optimization objective function is constructed. This function consists of two weighted sums: the first part is the output error cost. For each prediction step in the prediction time domain, the difference between the predicted output vector and the reference trajectory vector of that step is calculated first to obtain the error vector. The predicted output vector comes from the predicted state trajectory, and the reference trajectory vector comes from the target power allocation scheme. Next, a quadratic form operation is performed between the error vector and the adaptive weighting matrix. That is, the product of the error vector and the weighting matrix is ​​calculated first, and then the product is multiplied by the error vector to obtain the weighted squared error for this step. The weighted squared errors of all prediction steps are summed. The second part is the control increment cost. For each control step in the control time domain, the change in control input at the current moment is calculated first, that is, the control input of the current step minus the control input of the previous step. Then, a quadratic form operation is performed between this change and the control increment weighting matrix to obtain the control increment cost for this step. The control increment costs of all control steps are summed. The above two parts of the cost are summed to obtain the overall optimization objective function value. When solving this optimization problem, the state constraints must be satisfied: the state vector of each prediction step must be within the allowable range; the control constraints must be satisfied: each control input must be between the upper and lower limits; and the control increment constraints must be satisfied. A quadratic programming algorithm is used to find the control input sequence that minimizes the objective function under the premise of satisfying all constraints. This sequence contains the optimal control input value at each moment in the control time domain from the current moment, such as the fuel valve opening and excitation voltage, which is the first control sequence of the generator set.

[0045] S43. Perform rolling time-domain execution processing on the first control sequence to obtain the reference control input. In this step, the optimal control sequence of the generator set is a set of vectors arranged in chronological order, which contains multiple control vectors covering the control time domain length starting from the current moment. Each control vector contains multiple control components such as excitation reference voltage and speed regulation reference signal. Based on the rolling time-domain principle of model predictive control, only the first control vector in the sequence needs to be executed, while the subsequent control vectors are discarded because these subsequent vectors will be recalculated in the next sampling period to utilize the latest feedback information. Therefore, the first control vector in the optimal control sequence of the generator set is extracted and directly used as the reference control input of the generator set at the current moment. This reference control input contains various setpoints used to drive the actuators of the generator set, including excitation reference voltage, speed regulation reference signal, etc. These signals will be used as setpoints or feedforward inputs for the BP neural network adaptive PID controller in the subsequent step 4.4, providing the PID controller with the desired tracking target.

[0046] S44. Based on the reference control input, health status assessment results, and generator standardized dataset, adaptive PID parameters are obtained through online adaptive processing. In this step, the input layer vector is first constructed. The input features include key variables in the standardized operating dataset, the deviation between the reference control input and the actual output, and the fault severity and anomaly confidence in the health status assessment results. Then, the hidden layer calculation is performed: For each hidden layer node, the input values ​​of each node in the input layer are multiplied by the corresponding connection weights to obtain multiple products. All products are added together, and the bias of the hidden layer node is added to obtain the net input of the node. The net input is substituted into the hyperbolic tangent function, that is, the net input power of e is calculated minus the negative net input power of e, divided by the net input power of e plus the negative net input power of e, to obtain the output of the hidden layer node. Repeat this process to obtain the outputs of all hidden layer nodes, then proceed to the output layer for computation: the output layer contains three nodes, corresponding to K respectively. p T i T d For each output layer node, first multiply the output values ​​of all hidden layer nodes by their corresponding connection weights, sum all the products, and add the bias of the output layer node to obtain the net input of that node. Substitute this net input into the Sigmoid function to obtain a value between 0 and 1, and finally perform inverse normalization: For the proportional gain, first calculate the difference between the maximum and minimum proportional gain values, then multiply this difference by the output value of the first node in the output layer, and then add the minimum proportional gain value to obtain the actual proportional gain value; similarly, multiply the difference between the upper and lower limits of the integral time constant by the output of the second node plus the lower limit to obtain the integral time constant; use the difference between the upper and lower limits of the differential time constant... Multiplying the limit difference by the output of the third node plus the lower limit yields the differential time constant, and the final output of the generator set's adaptive PID parameters are these three values. This adaptive PID parameter tuning method adopts a three-layer feedforward neural network structure. The hidden layer uses the hyperbolic tangent function to ensure nonlinear mapping capability, and the output layer uses the Sigmoid function to constrain the parameters to the interval between 0 and 1, and then maps them to the actual parameter range through linear inverse normalization. This is the standard implementation method of neural network PID parameter self-tuning in the field of industrial control. Similar input layer designs are clearly described in many academic papers and patents on adaptive PID control, and will not be elaborated here.

[0047] S45. Based on the adaptive PID parameters, the target power allocation scheme, and the standardized generator set dataset, reverse error gradient descent processing is performed to obtain the second PID parameters; This step first involves determining the target power allocation scheme GST. PAS Obtain the reference input S at the current time. t Based on the standardized running dataset GS SOD Obtain the actual output SC S t Subtract S C Calculate the tracking error G z Meanwhile, the current control input DS T Subtract the control input DS from the previous time step T-1 Calculate the current control increment ΔD T Then define the performance index function X. N First calculate the tracking error G. z Square the value, then calculate the control increment ΔD. T Squared multiplied by the control increment ΔD T The penalty coefficient α is added together with the other two, and then multiplied by half to obtain the performance index value XN. Z During backpropagation, the error signal WC for the three nodes of the output layer, including the corresponding proportional gain, integral time constant, and derivative time constant, is calculated first. XH The error signal is equal to the tracking error G. z Multiply by the sign function sgn of the Jacobian information of the controlled object. The Jacobian information can be approximately obtained through system identification, which will not be elaborated here. Then multiply by the partial derivative PD of the output node's output with respect to the network output. a Then multiply by the derivative of the Sigmoid function, SD ID The output layer error signal WC is obtained. XH Then, the update amount ΔG of the output layer weight β. QZ Equals the learning rate ε multiplied by WC XH Then multiply by the output value YH of the hidden layer node SC ; For the hidden layer, first calculate the hidden layer error signal YH. WC The derivative YH of the activation function in the hidden layer DS Multiply by the output layer error signal WC XH The sum of the products of the output layer weights β and the hidden layer weights γ, the update amount ΔG. YHQ Equals the learning rate η multiplied by the hidden layer error signal YH WC Then multiply by the input value SR of the input layer node. C Add the corresponding update amount to the original weights of each layer to obtain the updated weights; re-execute the forward computation of S44 with the updated weights to obtain the new network output GX. SC Then, after inverse normalization, the values ​​between 0 and 1 in the output layer are mapped to the value range of the PID parameters, and finally the optimized second PID parameter OG of the generator set is output. SAPP .

[0048] S46. Perform instruction set synthesis processing on the second PID parameters, target power allocation scheme, and health status assessment results to obtain the control instruction set. In this step, firstly, the proportional gain, integral time constant, and derivative time constant specifically for the excitation control loop are extracted from the optimized adaptive PID parameters, and these values ​​are directly used as the generator set excitation control parameters. Secondly, for speed regulation control, the proportional gain, integral time constant, and derivative time constant of the speed regulation loop are also extracted from the optimized adaptive PID parameters. At the same time, the corresponding speed target value is calculated based on the active power target value in the target power allocation scheme. The above four values ​​are combined to form the generator set excitation control parameters. The generator set speed control parameters are then calculated, followed by the generator set protection setting values: Fault severity and anomaly confidence are extracted from the health status assessment results; the product of fault severity and anomaly confidence is first calculated, then multiplied by the health status attenuation coefficient to obtain an attenuation term; this attenuation term is subtracted from the value to obtain an adjustment coefficient; finally, this adjustment coefficient is multiplied by the baseline protection action threshold to obtain the dynamically adjusted protection action threshold, which decreases as the health status deteriorates, achieving early protection. Finally, the excitation control parameters, speed control parameters, and protection setting values ​​of each generator set are combined to form the complete control command for that generator set.

[0049] This invention uses a nonlinear dynamic model of the boiler, turbine, and generator to perform rolling time-domain prediction, anticipating load change trends and optimizing the control sequence. This solves the lag problem of large inertial systems, improves the speed of regulation and control, and significantly reduces dynamic deviation. In practice, for example, a gas turbine performs well with its PID parameters when running at full load, but when the load drops to 40%, the same parameters cause a speed overshoot of 12%, with fluctuations lasting for 20 seconds, seriously affecting the quality of power generation. In an uncommon special case where the unit experiences slight rotor imbalance, the fault severity is 0.4, and no alarm has been triggered. However, the vibration amplitude under fixed PID control gradually accumulates, and after 80 hours, the bearing wear intensifies, leading to a shutdown. This method can suppress vibration growth and solve the shutdown problem in advance by adjusting parameters based on the health status.

[0050] This invention adaptively adjusts the output error weighting matrix based on the health status, making the model more sensitive to the output errors of faulty generator sets. It proactively reduces the response speed to maintain stability. Simultaneously, the BP neural network adjusts the PID parameters online based on the health status assessment results. The model provides a macroscopically optimized control trajectory, and the BP neural network learns through backpropagation based on actual tracking errors. The synergy of these two aspects enables faulty generator sets to maintain stable operation even when the fault worsens. For example, if a generator set experiences a slow increase in leakage current due to insulation aging, and the fixed overcurrent protection threshold remains at 1.2 times the rated current, it will only trip when a phase-to-phase short circuit occurs, causing severe equipment damage. If the threshold can be adjusted according to the health status, the fault can be cleared earlier.

[0051] This invention dynamically reduces the protection action threshold based on the health status. For example, the baseline overcurrent protection threshold is 1.2 times the rated current. When the unit experiences severe vibration, the protection threshold automatically drops to 1.05 times the rated current, achieving early warning and protection and avoiding sudden damage. In simulation tests, this method significantly reduces the rate of unplanned downtime. In reality, for example, after a major overhaul of the unit, engineers need to spend about 8 hours retuning the PID parameters through trial and error. During this period, the unit cannot be connected to the grid, resulting in economic losses.

[0052] S5. Distributed effect feedback processing is performed based on the control instruction set to generate operation feedback data. Existing technologies lack the organic integration of distributed instruction routing, local actuator status awareness and closed-loop feedback, resulting in poor system reliability, no self-correction capability for execution deviations, and difficulty in adapting to dynamic changes, which seriously restricts the collaborative control efficiency of generator set clusters. To solve the above problems, the specific implementation steps are as follows: S51. Based on the control instruction set, the generator set node control instructions are obtained through distributed instruction processing. In this step, the coordination scheduling unit first broadcasts or sends the complete generator set control instruction set to all local intelligent control nodes through the communication network. For each generator set, after receiving the entire control instruction set, its corresponding local intelligent control node needs to extract the instruction set belonging to itself. The specific extraction process is as follows: the control instruction set is regarded as a set of instructions arranged in the order of generator set number. Each entry contains the excitation control parameters, speed control parameters and protection setting values ​​of the generator set. Based on the unit number, the corresponding position in the instruction set is located, and all parameters at that position are completely extracted to form an independent data packet. This data packet does not contain any information about other generator sets. Through the above extraction operation, each local intelligent control node obtains generator set node control instructions only for its own unit. This instruction contains all the set values ​​required for the unit to execute subsequent control, including the proportional gain, integral time constant, and derivative time constant of the excitation circuit, the proportional gain, integral time constant, derivative time constant, and target speed value of the speed regulation circuit, as well as the dynamically adjusted protection action threshold.

[0053] S52. Based on the generator set node control command, perform generator set control execution processing to obtain the generator set execution status. In this step, each local intelligent control node first obtains the generator set node control command at the current moment, including excitation control parameters, speed regulation control parameters, and protection setting values, as well as the actuator state vector at the current moment, such as the current fuel valve opening, current excitation current, and current steam valve position. Then, the target value in the node control command and the actuator state vector are input into the actuator transfer function for calculation. The transfer function processes each actuator channel separately: first, it calculates the deviation between the target value and the current actual value, and then multiplies the deviation by the corresponding proportional gain to obtain a proportional term. The deviation is integrated and multiplied by the integral gain to obtain the integral term; the deviation is differentiated and multiplied by the differential gain to obtain the differential term; the proportional term, integral term, and differential term are added together to obtain the preliminary control increment; this preliminary control increment is added to the output value of the previous moment to obtain the unlimited raw output value; then the raw output value is compared with the physical upper and lower limits of the actuator. If it exceeds the upper limit, the upper limit is taken; if it is lower than the lower limit, the lower limit is taken; otherwise, it remains unchanged to obtain the final actuator output value. The above calculation is repeated for all actuator channels to obtain the actuator output vector at that moment, which includes specific values ​​such as excitation current, fuel valve opening, and steam valve opening. After the actuator actually acts according to this output vector, the system records the actual action value of each actuator to form the generator set execution status.

[0054] S53. Perform execution effect feedback processing on the generator set execution status to generate operation feedback data. In this step, based on the information in the generator set execution status, the system first collects various actual operating parameters of the generator set in real time through sensors after the execution of control commands. The collection process includes: obtaining the actual value of generator set voltage from voltage transformer, obtaining the actual value of generator set current from current transformer, obtaining the actual value of generator set active power and reactive power from power transmitter, and obtaining the actual value of generator set frequency from frequency meter. These actual values ​​are all continuously changing analog quantities, which are converted into digital quantities after analog-to-digital conversion. Then, these collected actual values ​​are organized according to a preset data structure: first, the actual values ​​of active power, reactive power, frequency, voltage, and current are arranged in a fixed order to form a multi-dimensional vector. This vector does not contain any calculated intermediate quantities, but is all raw feedback quantities obtained directly from measurement. Finally, this vector is packaged into generator set operation feedback data.

[0055] This invention uses a distributed instruction routing and distribution method to send instruction sets to each generating unit via point-to-point or broadcast methods, avoiding the risk of all generating units going out of control due to a single point of communication failure in centralized scheduling. At the same time, each local node only extracts its own instructions, significantly reducing communication bandwidth usage. It generates precise actuator outputs through proportional, integral, and differential operations, and uses amplitude limiting processing to ensure that the actuators operate within a safe range. It collects the actual operating parameters of the generating units in real time through a closed-loop feedback monitoring method, and uses this data as the input for the next control cycle S1. The closed-loop feedback mechanism enables the system to cope with interference such as sudden changes in grid load and sensor noise, greatly reducing the frequency fluctuation range and significantly improving power generation quality and operational stability.

[0056] Example 2: Because existing technologies lack a mechanism that combines model prediction with neural network self-learning, they cannot optimize control parameters, leading to a decline in control quality over operating time and a lag in the response of the protection system. This severely restricts the safety and economy of generator sets. Please refer to [link to relevant documentation]. Figure 2 The diagram shown is a structural block diagram of a distributed control system for generator sets provided in this embodiment. The system includes a standardization module, a health assessment module, a power distribution module, a control command module, and an operation feedback module. The standardization module is used to collect the physical quantities of the target generator set and obtain a standardized dataset of the generator set through multi-channel standardization processing based on the physical quantities of the generator set. The health assessment module is used to perform spatiotemporal fault feature processing based on the generator set standardized dataset to obtain second-depth fault features, perform expectation-maximization confidence assessment on the second-depth fault features, and generate abnormal confidence and health status assessment results. The power allocation module is used to perform multi-source information reward processing on the standardized dataset of generator sets and the health status assessment results to obtain the second unit strategy action. Based on the health status assessment results and the second unit strategy action, the target power allocation scheme is obtained through topological consistency weighting. The control command module is used to obtain reference control input by dynamically executing the time domain based on the generator set standardized dataset and target power allocation scheme, and to perform online adaptive command set synthesis processing based on the reference control input to obtain the control command set; The runtime feedback module is used to perform distributed effect feedback processing based on the control instruction set and generate runtime feedback data.

[0057] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0058] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A distributed control method for generator sets, characterized in that, The steps of this method are as follows: collect the physical quantities of the target generator set, and obtain the generator set standardized dataset by performing multi-channel standardization processing based on the physical quantities of the generator set; Spatiotemporal fault feature processing is performed based on the standardized generator set dataset to obtain second-depth fault features. The second-depth fault features are then evaluated using expectation-maximization confidence to generate anomaly confidence and health status assessment results. Multi-source information reward processing is performed on the standardized dataset of generator sets and the health status assessment results to obtain the second unit strategy action. Based on the health status assessment results and the second unit strategy action, the target power allocation scheme is obtained through topology consistency weighting. Based on the standardized dataset of generator sets and the target power allocation scheme, the reference control input is obtained through dynamic time-domain execution processing. Based on the reference control input, the control command set is obtained through online adaptive instruction set synthesis processing. Distributed effect feedback processing is performed based on the control instruction set to generate runtime feedback data; The process involves dynamic time-domain processing based on the standardized generator set dataset and the target power allocation scheme, including: building a dynamic model based on the standardized generator set dataset and the target power allocation scheme to obtain the predicted state trajectory; The first control sequence is obtained by performing constraint control optimization based on the predicted state trajectory, the target power allocation scheme, and the health status assessment results. The first control sequence is subjected to rolling time-domain execution processing to obtain the reference control input.

2. The distributed control method for generator sets according to claim 1, characterized in that, Based on the physical quantities of the generator set, a multi-channel standardization process is performed, including: The original signals from the generator set sensors are obtained by synchronously processing the physical quantities of the generator set through multiple channels. Adaptive anomaly detection and correction processing is performed on the original signals from the generator set sensors to generate generator set cleaning signals. The generator set cleaning signal is standardized to obtain a standardized generator set dataset.

3. The distributed control method for generator sets according to claim 1, characterized in that, Spatiotemporal fault feature processing based on standardized generator set datasets includes: The first deep fault features are obtained by extracting spatiotemporal features from the standardized generator set dataset; Based on the first-depth fault features, a preliminary fault type determination is performed, and a preliminary fault type determination result is generated. Deep fault feature processing is performed on the standardized dataset of generator sets and the preliminary fault type discrimination results to obtain the second deep fault features.

4. The distributed control method for generator sets according to claim 1, characterized in that, The expected-maximization confidence assessment of the characteristics of second-depth faults includes: Based on the characteristics of the second-depth fault, a Gaussian mixture model parameter set is generated through expectation maximization. Based on the parameter set of the Gaussian mixture model and the second-depth fault features, the fault type and severity are determined to obtain the fault type and fault severity. Anomaly confidence assessments are performed on fault types, Gaussian mixture model parameter sets, and second-depth fault features to generate anomaly confidence and health status assessment results.

5. The distributed control method for generator sets according to claim 1, characterized in that, Multi-source information reward processing is applied to the standardized generator set dataset and health status assessment results, including: The model observation status is obtained through multi-source information fusion processing based on the standardized generator set dataset and health status assessment results; Based on the model observation state, a deep-determined gradient processing is performed to obtain the first unit's strategy action; The strategy actions of the first unit, the standardized dataset of the generator unit, and the health status assessment results are processed with multi-objective compliance rewards to obtain the strategy actions of the second unit.

6. The distributed control method for generator sets according to claim 1, characterized in that, Based on the health status assessment results and the second unit's strategy actions, a topology-consistent weighted processing method is used, including: Topological consistency processing is performed based on the standardized generator set dataset and health status assessment results to generate consistent state variables; Distributed Newton optimization is performed based on consistent state variables to obtain the Newton optimization allocation solution. An adaptive weighted fusion of the second unit's strategy actions and the Newton-optimal allocation solution is performed to obtain the target power allocation scheme.

7. The distributed control method for generator sets according to claim 1, characterized in that, Online adaptive instruction set synthesis based on reference control input includes: Adaptive PID parameters are obtained through online adaptive processing based on the reference control input, health status assessment results, and generator set standardized dataset; The second PID parameters are obtained by performing back error gradient descent processing based on adaptive PID parameters, target power allocation scheme and generator set standardized dataset; The second PID parameters, the target power allocation scheme, and the health status assessment results are processed by instruction set integration to obtain the control instruction set.

8. The distributed control method for generator sets according to claim 1, characterized in that, Distributed effect feedback processing based on control instruction sets includes: Based on the control instruction set, distributed instruction processing is used to obtain the generator set node control instructions; Based on the generator set node control commands, the generator set control execution process is performed to obtain the generator set execution status; The execution status of the generator set is processed to provide feedback on the execution effect, and operational feedback data is generated.

9. The distributed control method for generator sets according to claim 1, characterized in that, The physical quantities of a generator set include generator set vibration acceleration, generator set temperature, generator set current, generator set voltage, generator set frequency, generator set active power, generator set reactive power, and generator set speed.

10. A system applied to the distributed control method for generator sets according to any one of claims 1-9, characterized in that, The system includes: The standardization module is used to collect the physical quantities of the target generator set and obtain a standardized dataset of the generator set through multi-channel standardization processing based on the physical quantities of the generator set. The health assessment module is used to perform spatiotemporal fault feature processing based on the generator set standardized dataset to obtain second-depth fault features, perform expectation-maximization confidence assessment on the second-depth fault features, and generate abnormal confidence and health status assessment results. The power allocation module is used to perform multi-source information reward processing on the standardized dataset of generator sets and the health status assessment results to obtain the second unit strategy action. Based on the health status assessment results and the second unit strategy action, the target power allocation scheme is obtained through topological consistency weighting. The control command module is used to obtain reference control input by dynamically executing the time domain based on the generator set standardized dataset and target power allocation scheme, and to perform online adaptive command set synthesis processing based on the reference control input to obtain the control command set; The runtime feedback module is used to perform distributed effect feedback processing based on the control instruction set and generate runtime feedback data.

Images