Load optimization control method and system for generator set

Abstract

The invention discloses a load optimization control method and system for a generator set, belongs to the technical field of power systems, and solves the problem that complex dynamic changes are difficult to capture when a set load is calculated through a difference value between a theoretical load and a real load in an existing method. The method comprises the steps of selecting operation correlation parameters based on a feature priority strategy, constructing a load prediction model based on a long short-term memory network, identifying and predicting operation load data, outputting a load rate prediction threshold interval of a generator set, and calculating a load distribution value corresponding to the generator set in a prediction period based on an optimal distribution function. According to the method, the operation correlation parameters are selected based on the feature priority strategy to obtain the correlation feature set which is high in correlation degree with different types of generator sets and has large influence on the load rates of the generator sets, the adaptability of the load prediction model to the different types of generator sets is improved, and the accuracy of adjusting the operation loads of the generator sets is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and particularly relates to a load optimization control method and system for a generator set. Background Art

[0002] With the continuous growth of energy demand and the increasing complexity of power systems, the load control of generator sets plays a crucial role in the operation of power systems. Traditional load control methods for generator sets mostly adopt simple average distribution or experience-based scheduling methods. Although they can meet the load demand to a certain extent, when facing rapid load changes, the traditional methods have a slow response speed and are prone to lag phenomena, affecting the stability and reliability of power systems.

[0003] Chinese Patent CN119393231A discloses a load control method and device for a gas turbine generator set. The method includes the following steps: obtaining the load control parameters of the gas turbine generator set in the current period, and obtaining the theoretical load of the gas turbine generator set in the current period according to the load control parameters; obtaining the actual load of the gas turbine generator set in the current period; calculating the set load of the gas turbine generator set in the next period according to the theoretical load and the actual load of the gas turbine generator set in the current period; in the next period, starting from the start time of the next period, using the set load as the reference value of the PID control system to control the gas turbine generator set; however, the existing method calculates the set load through the difference between the theoretical load and the actual load, which is difficult to capture complex dynamic changes and has limited adaptability to different types of generator sets. In view of the above problems, we propose a load optimization control method and system for a generator set. Summary of the Invention

[0004] The purpose of the present invention is to provide a load optimization control method and system for a generator set in view of the deficiencies of the prior art, and solve the problems that the existing method calculates the set load through the difference between the theoretical load and the actual load, which is difficult to capture complex dynamic changes and has limited adaptability to different types of generator sets.

[0005] The present invention is implemented as follows. A load optimization control method for a generator set, the load optimization control method for a generator set includes: Obtaining the day-ahead operation correlation parameters of the generator set, selecting the operation correlation parameters based on the feature priority strategy to obtain a correlation feature set, constructing a load prediction model based on a long short-term memory network, and iteratively training the load prediction model using the correlation feature set; Collect the operation load data of the generator set in real time, normalize the operation load data, and the load prediction model identifies and predicts the operation load data, and outputs the load rate prediction threshold interval of the generator set; Load the load rate prediction threshold interval and the associated feature set, and perform fusion processing on the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm to obtain the feature fusion set; Construct an optimization allocation function with the operation cost of the unit, the success rate of primary frequency modulation, and the response time as constraints, and calculate the load allocation value corresponding to the generator set within the prediction period based on the optimization allocation function combined with the feature fusion set, and trigger the load allocation instruction; In response to the load allocation instruction, the PID control module adjusts the operation load of the generator set based on the load allocation instruction.

[0006] Preferably, the method for selecting operation correlation parameters based on the feature priority strategy includes: Load the operation correlation parameters, where the operation correlation parameters include unit type, active power, reactive power, load rate, load type, grid frequency, frequency fluctuation, environmental temperature, environmental humidity, wind force, altitude, unit rated power, power factor, fuel consumption, equipment failure rate, equipment risk, and number of shutdowns. Construct a two-dimensional feature attribute system based on the operation correlation parameters. The two-dimensional feature attribute system consists of three types of attributes based on the time dimension and three types of attributes based on the load rate. The two-dimensional feature attribute system comprehensively divides the timeliness of the parameters, and is successively a static fact attribute matrix, a dynamic influence attribute matrix, and an associated interaction attribute matrix; Grab and distinguish the operation correlation parameters based on the two-dimensional feature attribute system, fill the static fact attribute matrix, the dynamic influence attribute matrix, and the associated interaction attribute matrix, and assign weights to the operation correlation parameters in the static fact attribute matrix and the dynamic influence attribute matrix respectively based on the principal component analysis method; Load the static fact attribute matrix and the dynamic influence attribute matrix after weighting the operation correlation parameters, multiply the static fact attribute matrix and the dynamic influence attribute matrix by the associated interaction attribute matrix respectively to obtain the static association matrix and the dynamic association matrix; Load the static association matrix and the dynamic association matrix, accumulate the static association matrix and the dynamic association matrix based on Pandas to obtain the attribute association matrix, and calculate the correlation coefficient between the operation correlation parameters and the load rate based on the Pearson algorithm; Define the parameter determination threshold by combining the static association matrix, the dynamic association matrix and the comprehensive feature priority strategy; Load the correlation coefficient of the operation correlation parameters, and select the operation correlation parameters with a correlation coefficient greater than the parameter determination threshold, and integrate to obtain the associated feature set.

[0007] Preferably, the correlation coefficient between the operation correlation parameters and the load rate is calculated by the following formula: (1) Among them, represents the correlation coefficient between the operation correlation parameter and the load rate, are respectively the attribute correlation matrix the load rate of the correlation parameter and the eigenvalue of the correlation parameter in it, is the mean value of the attribute correlation matrix, are respectively the weight values of the correlation parameter characteristics in the static correlation matrix and the dynamic correlation matrix; The definition calculation formula of the parameter determination threshold is as follows: (2) (3) Among them, represents the parameter determination threshold, are respectively the static correlation matrix and the mean value of the static correlation matrix, are respectively the dynamic correlation matrix and the mean value of the dynamic correlation matrix, respectively represent the regression coefficients of the static correlation matrix and the dynamic correlation matrix, is the minimum value of the sum of the static correlation matrix and the dynamic correlation matrix superimposed with the regression coefficient, represents the number of correlation parameter characteristics, represents the regularization parameter of the feature priority strategy, are respectively the maximum value, minimum value, and mean value of the dynamic correlation matrix weight, are respectively the maximum value, minimum value, and mean value of the static correlation matrix weight.

[0008] Preferably, the method for iteratively training the load prediction model using the correlation feature set includes: Load the correlation feature set, process the correlation feature set using the time-domain simulation method, and standardize the correlation feature set using the maximum-minimization method, and divide the correlation feature set into a training set and a test set; Load the pre-constructed load prediction model, determine the initial values of the hyperparameters of the load prediction model based on the improved vulture search-expectation maximization algorithm, set the transfer function between the input gate and the Seq2seq model, the transfer function between the Seq2seq model and the Gaussian mixture model, and set the loss function of the load prediction model; Obtain the training set, set the learning rate of the load prediction model to 0.002, iteratively train the load prediction model with the training set for 100 rounds, and output the initial rough model; Load the initial coarse model, gradually eliminate the prediction results of the initial coarse model through a multi-step inverse diffusion process, set the learning rate of the load prediction model to 0.0002, and use the training set and the prediction results of the initial coarse model to iteratively train the load prediction model for 50 rounds to output the initial fine model. During model training, use the Adam gradient optimization algorithm to optimize the Seq2seq model and use the variational Bayesian method to optimize the Gaussian mixture model. Iteratively update the hyperparameters of the load prediction model until the loss function converges to obtain the trained load prediction model; Obtain the test set, and use the method of layer-by-layer search to input the test set into the trained load prediction model for layer-by-layer separate testing, output the test results, and determine whether the accuracy rate of the test results and the true results meets the preset accuracy threshold; If the accuracy rate of the test results and the true results meets the preset accuracy threshold, output the converged load prediction model.

[0009] Preferably, the load prediction model uses a long short-term memory network as the initial model. The initial model includes an input gate, a forget gate, and an output gate connected in sequence. When constructing the load prediction model, use the Seq2seq model to replace the forget gate of the initial model. The Seq2seq model includes an encoder and a decoder. Freeze the decoder of the Seq2seq model, use the spatio-temporal neural network architecture to replace the decoder, and introduce the dynamic time warping algorithm into the spatio-temporal neural network architecture. The spatio-temporal neural network architecture combined with the dynamic time warping algorithm is used to normalize the operating load data and extract the associated features associated with the operating load data. Introduce a Gaussian mixture model based on the Stackelberg game algorithm between the Seq2seq model and the output gate. The Gaussian mixture model is used to use the load rate as the leader and the associated features as the follower, and combine the kernel density estimation method to determine the load rate prediction threshold interval.

[0010] Preferably, the method for fusing and processing the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm includes: Load the load rate prediction threshold interval, perform differential processing on the load rate prediction threshold interval to obtain the load rate difference sequence of the unit's time series; Determine the operating load of the generator set based on the correlation coefficient of the load rate difference sequence and the operating correlation parameters in the associated feature set; (4) (5) Wherein, is the operating load of the generator set at time is the rated active power of the generator set, is the mean value of the load rate prediction threshold interval at time represent the mean of the predicted threshold interval of the moment load rate, represent the difference value of the load rate of the generating unit at the moment, are respectively the maximum value of the correlation coefficient of the operation-related parameters and the correlation coefficient, represent the number of associated parameter features; Adopt the aggregated equivalent method to compress the difference value of the generating unit node, take the minimum cost as the constraint, and perform weighted summation on the operation load of the generating unit and the associated feature set to construct a feature fusion set; Among them, the feature fusion set is expressed as: (6) Among them, represent the output representation of the feature fusion set, is the number of generating units, respectively represent the generating unit photovoltaic abandonment component coefficient and fuel cost coefficient, represent the number of nodes for difference compression of the generating unit node, represent the generating unit threshold effect coefficient, respectively represent the maximum and minimum values of the active power output of the generating unit.

[0011] Preferably, when constructing the optimization distribution function with the operation cost of the unit, the success rate of primary frequency modulation, and the response time as constraints, the optimization distribution function is expressed as: (7) Among them, represent with the operation cost of the unit, the success rate of primary frequency modulation, and the response time as constraints, the load distribution value allocated by the generating unit based on the optimization distribution function at the moment, are respectively the operation cost of the unit, the frequency modulation scheduling cost, and the response cost, are respectively the primary frequency modulation constraint coefficient and the response time constraint coefficient.

[0012] Another method, the present invention also provides a load optimization control system for a generating unit, and the load optimization control system for the generating unit includes: A feature optimization module, configured to obtain the operation-related parameters of the generating unit on the previous day, select the operation-related parameters based on the feature priority strategy to obtain an associated feature set, construct a load prediction model based on a long short-term memory network, and iteratively train the load prediction model with the associated feature set; An operating data acquisition module, which is used to collect the operating load data of the generator set in real time, normalize the operating load data, and the load prediction model identifies and predicts the operating load data, and outputs the load rate prediction threshold interval of the generator set; A feature fusion module, which loads the load rate prediction threshold interval and the associated feature set, and performs fusion processing on the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm to obtain a feature fusion set; An optimization control module, which is used to construct an optimization allocation function with the operating cost of the unit, the success rate of primary frequency modulation, and the response time as constraints, calculate the load allocation value corresponding to the generator set within the prediction period based on the optimization allocation function combined with the feature fusion set, trigger a load allocation instruction, and in response to the load allocation instruction, the PID control module adjusts the operating load of the generator set based on the load allocation instruction.

[0013] Preferably, the feature optimization module includes: A parameter grabbing unit, which is used to obtain the operating associated parameters of the generator set on the previous day; A feature optimization unit, which selects the operating associated parameters based on the feature priority strategy to obtain an associated feature set; A load prediction unit, which constructs a load prediction model based on a long short-term memory network, iteratively trains the load prediction model using the associated feature set, and outputs a converged load prediction model.

[0014] Compared with the prior art, the embodiments of the present application mainly have the following beneficial effects: In the embodiments of the present invention, an associated feature set with a high degree of association with different types of generator sets and a great influence on the load rate of the generator set is selected based on the feature priority strategy, thereby ensuring the accuracy of the load rate prediction of the generator set by the load prediction model and the robustness of the model, improving the adaptability of the load prediction model to different types of generator sets, and ensuring the accuracy of adjusting the operating load of the generator set. It overcomes the problem that the existing method calculates and sets the load by the difference between the theoretical load and the actual load, which is difficult to capture complex dynamic changes and has limited adaptability to different types of generator sets.

[0015] In the embodiments of the present invention, the feature priority strategy method can systematically screen out a feature set highly correlated with the load rate from a large number of operation-related parameters. The two-dimensional feature attribute system classifies parameters according to the time and load rate dimensions, which helps to clearly identify the impact of different parameters on the load rate, and improves the logic and systematicness of feature processing. The division of the static fact attribute matrix, dynamic influence attribute matrix, and associated interaction attribute matrix can distinguish the timeliness and interactivity of parameters, providing a basis for subsequent feature weight allocation. Moreover, through scientific algorithms and methods such as PCA and Pearson correlation coefficient, the objectivity, accuracy, and adaptability of feature selection are ensured. The high-quality associated feature set can significantly improve the accuracy and robustness of the load prediction model, providing reliable data support for subsequent optimization control.

[0016] In the embodiments of the present invention, a load prediction model and a training method are provided. The load prediction model uses a long short-term memory network as the initial model, and introduces a Seq2seq model, a spatio-temporal neural network architecture, a dynamic time warping algorithm, and a Gaussian mixture model based on the Stackelberg game algorithm. The Seq2seq model can achieve a direct mapping from the input sequence to the output sequence without complex feature engineering. The DTW algorithm can perform dynamic alignment on time series data, further enhancing the model's sensitivity to time series changes. The Gaussian mixture model based on the Stackelberg game algorithm can model the distribution of the load rate. Combining with the Stackelberg game algorithm, it can dynamically adjust the model parameters to improve the prediction accuracy. At the same time, the load prediction model can identify and predict the operating load data, and output a load rate prediction threshold interval, overcoming the problem that the point prediction in the prior art cannot effectively reflect the uncertainty in load prediction and is difficult to provide reliable information on the prediction results. Through the prediction threshold interval, the uncertainty of load prediction can be quantified, providing more comprehensive information for the operation of the power system and a more flexible decision-making basis for the power system dispatching, which helps to formulate a more robust operation strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flowchart of the implementation of the load optimization control method for a generator set provided by the present invention.

[0018] Figure 2 It shows a schematic flowchart of the implementation of the method for selecting operation-related parameters based on the feature priority strategy.

[0019] Figure 3 It shows a schematic flowchart of the implementation of the method for iteratively training the load prediction model using the associated feature set.

[0020] Figure 4 It shows a schematic flowchart of the implementation of the method for fusing and processing the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm.

[0021] Figure 5 Shows a schematic structural diagram of a load optimization control system for a generator set. Specific implementation manners

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs; the terms used in the description of this application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the description and claims of this application and the above drawings are intended to cover non-exclusive inclusion. The terms "first", "second", etc. in the description and claims of this application or the above drawings are used to distinguish different objects and not to describe a specific order.

[0023] Existing methods calculate the set load through the difference between the theoretical load and the actual load, which is difficult to capture complex dynamic changes and has limited adaptability to different types of generator sets. To address the above problems, we propose a load optimization control method and system for generator sets. Briefly, when the method is implemented, first, the operation-related parameters of the generator set for the day ahead are obtained, the operation-related parameters are selected based on the feature priority strategy to obtain an associated feature set, and at the same time, a load prediction model based on a long short-term memory network is constructed. Then, the operation load data of the generator set is collected in real time, and the operation load data is normalized. The load prediction model identifies and predicts the operation load data, and outputs the predicted threshold interval of the load rate of the generator set. Then, based on the differential feature extraction algorithm, the predicted threshold interval of the load rate and the associated feature set are fused to obtain a feature fusion set. Finally, an optimization distribution function is constructed with the operation cost of the unit, the success rate of primary frequency regulation, and the response time as constraints, and the load distribution value corresponding to the generator set in the prediction period is calculated based on the optimization distribution function combined with the feature fusion set. In the embodiments of the present invention, the operation-related parameters are selected based on the feature priority strategy to obtain an associated feature set with a high degree of association with different types of generator sets and a great influence on the load rate of the generator set, thereby ensuring the accuracy of the load prediction model for predicting the load rate of the generator set and the robustness of the model, improving the adaptability of the load prediction model to different types of generator sets, and ensuring the accuracy of adjusting the operation load of the generator set. It overcomes the problems of existing methods that calculate the set load through the difference between the theoretical load and the actual load, which is difficult to capture complex dynamic changes and has limited adaptability to different types of generator sets.

[0024] The embodiments of the present invention provide a load optimization control method for a generator set, Figure 1 Shows a schematic implementation flow diagram of a load optimization control method for a generator set. The load optimization control method for a generator set includes: Step S10: Obtain the operation-related parameters of the generator set on a daily basis, select the operation-related parameters based on the feature priority strategy, and screen the operation-related parameters using the feature priority strategy. This can accurately extract the feature set highly correlated with the load rate of the generator set. This strategy can effectively remove redundant features, reduce the model complexity, improve the training efficiency and prediction accuracy of the model at the same time. Moreover, through the combination of the feature priority strategy and the LSTM model, it can effectively cope with the complex operation conditions of different types of generator sets, improve the adaptability of the model, obtain the associated feature set, construct a load prediction model based on the long short-term memory network, and iteratively train the load prediction model using the associated feature set; It should be noted that the generator set includes but is not limited to unit voltage, current, frequency, power, power factor, load rate, operation time, ambient temperature, ambient humidity, altitude, and the generator set includes but is not limited to diesel generator sets, gasoline generator sets, gas generator sets, wind turbine generator sets, solar generator sets, hydraulic generator sets, coal-fired generator sets.

[0025] Step S20: Real-time collect the operation load data of the generator set, and perform normalization processing on the operation load data. Real-time collection of the operation load data of the generator set can timely capture the dynamic changes of the load and provide the latest data support for subsequent prediction and control. The load prediction model identifies and predicts the operation load data, and outputs the load rate prediction threshold interval of the generator set; Step S30: Load the load rate prediction threshold interval and the associated feature set, and perform fusion processing on the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm to obtain the feature fusion set. Performing fusion processing on the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm can further explore the dynamic features of the load change; Step S40: Construct an optimization allocation function with the operation cost of the unit, the success rate of primary frequency regulation, and the response time as constraints, and calculate the corresponding load allocation value of the generator set during the prediction period based on the optimization allocation function combined with the feature fusion set, and trigger the load allocation instruction; Step S50: In response to the load allocation instruction, the PID control module adjusts the operation load of the generator set based on the load allocation instruction. The PID control module can be flexibly adjusted according to different load allocation instructions to adapt to the operation requirements of different types of generator sets. The introduction of the PID control module can effectively cope with the real-time requirements of load changes and ensure the stable operation of the generator set.

[0026] In the embodiments of the present invention, based on the feature priority strategy, operation-related parameters are selected to obtain an association feature set with a high degree of association with different types of generator sets and a great influence on the load factor of the generator sets. Thus, the accuracy of the load prediction model for predicting the load factor of the generator sets and the robustness of the model are ensured, the adaptability of the load prediction model to different types of generator sets is improved, and the accuracy of adjusting the operating load of the generator sets is ensured. It overcomes the problem that the existing method calculates and sets the load through the difference between the theoretical load and the actual load, which is difficult to capture complex dynamic changes and has limited adaptability to different types of generator sets.

[0027] The embodiments of the present invention provide a method for selecting operation-related parameters based on the feature priority strategy. Figure 2 The schematic diagram of the implementation process of the method for selecting operation-related parameters based on the feature priority strategy is shown. The method for selecting operation-related parameters based on the feature priority strategy includes: Step S101, load operation-related parameters. Among them, the operation-related parameters include, but are not limited to, unit type, active power, reactive power, load factor (the load factor refers to the ratio of the actual output power of the generator to the rated power, which is used to measure the load utilization rate and operating efficiency of the generator), load type (loads can be divided into resistive, inductive, and capacitive. Different types of loads have different effects on the power demand and voltage stability of the generator), grid frequency, frequency fluctuation, environmental temperature (environmental factors such as temperature, humidity, and altitude will also affect the load-bearing capacity of the generator), environmental humidity, wind force, altitude, unit rated power, power factor, fuel consumption, equipment failure rate, equipment risk, number of shutdowns. Based on the operation-related parameters, a two-dimensional feature attribute system is constructed. The two-dimensional feature attribute system consists of three types of attributes based on the time dimension and three types of attributes based on the load factor. The two-dimensional feature attribute system divides the comprehensive parameters according to timeliness, and in turn is a static fact attribute matrix, a dynamic influence attribute matrix, and an association interaction attribute matrix; It should be noted that the operation-related parameters cover various aspects of information such as unit type, power, load factor, environmental conditions, and equipment status, ensuring the comprehensiveness of feature selection. The two-dimensional feature attribute system classifies the parameters according to the time and load factor dimensions, which helps to clearly identify the influence of different parameters on the load factor, improves the logic and systematicness of feature processing. The division of the static fact attribute matrix, dynamic influence attribute matrix, and association interaction attribute matrix can distinguish the timeliness and interactivity of the parameters, providing a basis for subsequent feature weight assignment.

[0028] Step S102: Based on the two-dimensional feature attribute system, distinguish and capture the operation-related parameters, fill the static fact attribute matrix, dynamic influence attribute matrix, and associated interaction attribute matrix, and assign weights to the operation-related parameters in the static fact attribute matrix and dynamic influence attribute matrix respectively based on the principal component analysis method; through weight assignment, the parameters with greater influence on the load rate can be highlighted, providing a scientific basis for subsequent feature selection.

[0029] Step S103: Load the static fact attribute matrix and dynamic influence attribute matrix after weight assignment of the operation-related parameters, multiply the static fact attribute matrix and dynamic influence attribute matrix with the associated interaction attribute matrix respectively to obtain the static association matrix and dynamic association matrix. Among them, the generation of the static association matrix and dynamic association matrix can more comprehensively reflect the direct and indirect associations between the parameters and the load rate, providing richer information for subsequent feature fusion. The introduction of the dynamic association matrix can adapt to the dynamic characteristics of load changes and enhance the model's response ability to real-time data.

[0030] Step S104: Load the static association matrix and dynamic association matrix, accumulate the static association matrix and dynamic association matrix based on Pandas to obtain the attribute association matrix, and calculate the correlation coefficient between the operation-related parameters and the load rate based on the Pearson algorithm; The correlation coefficient between the operation-related parameters and the load rate is calculated by the following formula: (1) where, represents the correlation coefficient between the operation-related parameters and the load rate, are the load rate of the associated parameter and the eigenvalue of the associated parameter in the attribute association matrix respectively, is the mean value of the attribute association matrix, are the weight values of the associated parameter features in the static association matrix and dynamic association matrix respectively.

[0031] It should be noted that by calculating the correlation coefficient between the operation-related parameters and the load rate, the influence degree of each parameter on the load rate can be quantified. The closer the absolute value of the correlation coefficient is to 1, the stronger the linear relationship between the parameter and the load rate. Selecting the parameters with a correlation coefficient greater than the set threshold can ensure that the selected feature set makes a significant contribution to the load rate prediction. The calculation of the correlation coefficient and the setting of the threshold can be adjusted according to different types of generator sets and operating conditions, enhancing the adaptability and flexibility of the feature selection method. The feature selection method based on the correlation coefficient is data-driven, can objectively reflect the relationship between the parameters and the load rate, and avoids the uncertainty brought by subjective judgment.

[0032] In the embodiments of the present invention, the Pearson correlation coefficient can quantify the linear relationship between parameters and the load rate, providing a clear quantitative index for feature selection. Moreover, calculating the correlation coefficient based on data ensures the objectivity and data-driven nature of feature selection, avoiding the uncertainty of subjective judgment.

[0033] Step S105: Define a parameter determination threshold by combining the static correlation matrix, the dynamic correlation matrix, and the comprehensive feature priority strategy. The definition calculation formula of the parameter determination threshold is as follows: (2) (3) Wherein, represents the parameter determination threshold, are respectively the static correlation matrix and the mean value of the static correlation matrix, are respectively the dynamic correlation matrix and the mean value of the dynamic correlation matrix, respectively represent the regression coefficients of the static correlation matrix and the dynamic correlation matrix. The regression coefficient reflects the degree of linear influence of each feature on the load rate. By calculating the norm of the regression coefficient, the importance of the feature can be quantified. The regression coefficients of the static correlation matrix and the dynamic correlation matrix can be 0.5 - 0.8. is the minimum value of the sum of the static correlation matrix and the dynamic correlation matrix superimposed with the regression coefficient, represents the number of associated parameter features, represents the regularization parameter of the feature priority strategy, are respectively the maximum value, the minimum value, and the mean value of the weights of the dynamic correlation matrix, are respectively the maximum value, the minimum value, and the mean value of the weights of the static correlation matrix. In the embodiments of the present invention, by adjusting the regularization parameter , the strictness of feature selection can be dynamically adjusted according to specific application scenarios, enhancing the flexibility of the model. Moreover, the formula combines the regression coefficient and the regularization parameter, which can scientifically define the parameter determination threshold and ensure the objectivity and effectiveness of feature selection.

[0034] It should be noted that the definition of the threshold is based on the comprehensive feature priority strategy, which can adapt to different types of generator sets and operating conditions, enhancing the flexibility and universality of the method. Step S106: Load the correlation coefficients of the running associated parameters, select the running associated parameters whose correlation coefficients are greater than the parameter determination threshold, and integrate them to obtain an associated feature set. A high-quality associated feature set can significantly improve the accuracy and robustness of the load prediction model, providing reliable data support for subsequent optimal control.

[0035] In the embodiments of the present invention, the feature priority strategy method can systematically screen out a feature set highly correlated with the load rate from a large number of operation-related parameters. The two-dimensional feature attribute system classifies the parameters according to the time and load rate dimensions, which helps to clearly identify the influence of different parameters on the load rate, and improves the logic and systematicness of feature processing. The division of the static fact attribute matrix, dynamic influence attribute matrix, and correlation interaction attribute matrix can distinguish the timeliness and interaction of parameters, providing a basis for subsequent feature weight allocation. Moreover, through scientific algorithms and methods such as PCA and Pearson correlation coefficient, the objectivity, accuracy, and adaptability of feature selection are ensured. The high-quality correlation feature set can significantly improve the accuracy and robustness of the load prediction model, providing reliable data support for subsequent optimization control.

[0036] The embodiments of the present invention provide a method for iteratively training a load prediction model using a correlation feature set. Figure 3 The schematic diagram of the implementation process of the method for iteratively training a load prediction model using a correlation feature set is shown. The method for iteratively training a load prediction model using a correlation feature set includes: Step S201, load the correlation feature set, process the correlation feature set using the time-domain simulation method, and standardize the correlation feature set using the maximum-minimum method, and divide the correlation feature set into a training set and a test set; In the embodiments of the present invention, processing the correlation feature set using the time-domain simulation method can effectively eliminate noise and outliers, improve data quality, and the allocation ratio of the training set and the test set can be 3 - 4:1.

[0037] Step S202, load the pre-constructed load prediction model, determine the initial values of the hyperparameters of the load prediction model based on the improved vulture search - expectation maximization algorithm, set the transfer function between the input gate and the Seq2seq model, and the transfer function between the Seq2seq model and the Gaussian mixture model. In this embodiment, the transfer function can be Rule, Sigmoid function, and set the loss function of the load prediction model. The loss function of the model can be the cross-entropy loss function. Setting the transfer function between the input gate and the Seq2seq model, and the transfer function between the Seq2seq model and the Gaussian mixture model can achieve effective connection between different modules and improve the overall performance of the model; In the embodiments of the present invention, determining the initial values of the hyperparameters of the load prediction model based on the improved vulture search - expectation maximization algorithm can effectively avoid the instability and slow convergence speed caused by random initialization in traditional methods, and clearly setting the transfer function and loss function of each part provides a clear framework and goal for subsequent model training.

[0038] Step S203: Obtain the training set, set the learning rate of the load prediction model to 0.002, and iteratively train the load prediction model with the training set for 100 rounds to output the initial rough model; Step S204: Load the initial rough model, gradually eliminate the prediction results of the initial rough model through a multi-step inverse diffusion process, set the learning rate of the load prediction model to 0.0002, and iteratively train the load prediction model with the training set and the prediction results of the initial rough model for 50 rounds to output the initial fine model. During model training, use the Adam gradient optimization algorithm to optimize the Seq2seq model and use the variational Bayesian method to optimize the Gaussian mixture model. Iteratively update the hyperparameters of the load prediction model until the loss function converges to obtain the trained load prediction model. In this embodiment, the loss function convergence index can be set to 0.08 - 0.1. By gradually eliminating the noise in the prediction results of the initial rough model through the multi-step inverse diffusion process, the prediction accuracy of the model can be further improved. The inverse diffusion process can gradually optimize the output of the model and enhance the adaptability of the model to complex data. Using the Adam gradient optimization algorithm to optimize the Seq2seq model can combine the adaptive learning rate and the momentum term to accelerate the convergence speed of the model. Using the variational Bayesian method to optimize the Gaussian mixture model can enhance the generalization ability and robustness of the model; Step S205: Obtain the test set, and use the layer-by-layer search method to input the test set into the trained load prediction model for layer-by-layer separate testing to output the test results. Using the layer-by-layer search method to input the test set into the trained load prediction model for layer-by-layer separate testing can comprehensively evaluate the performance of the model at different levels. And the layer-by-layer search method can discover possible problems of the model at certain levels and provide a basis for subsequent optimization; Step S206: Determine whether the accuracy rate of the test results and the true results meets the preset accuracy threshold. In this embodiment, the accuracy threshold can be set to 0.85 - 0.95; Step S207: If the accuracy rate of the test results and the true results meets the preset accuracy threshold, output the converged load prediction model; If the accuracy rate of the test results and the true results does not meet the preset accuracy threshold, return to Step S203 and continue to iteratively train the model.

[0039] In this embodiment, the load prediction model uses a long short-term memory network as the initial model. The initial model includes an input gate, a forget gate, and an output gate connected in sequence. When constructing the load prediction model, a Seq2seq model is used to replace the forget gate of the initial model. The Seq2seq model includes an encoder and a decoder. The decoder of the Seq2seq model is frozen, and a spatio-temporal neural network architecture is used to replace the decoder. A dynamic time warping algorithm is introduced into the spatio-temporal neural network architecture. The spatio-temporal neural network architecture combined with the dynamic time warping algorithm is used to normalize the operating load data and extract the associated features associated with the operating load data. A Gaussian mixture model based on the Stackelberg game algorithm is introduced between the Seq2seq model and the output gate. The Gaussian mixture model is used to take the load rate as the leader and the associated features as the follower, and combine the kernel density estimation method to determine the load rate prediction threshold interval.

[0040] In the embodiment of the present invention, a load prediction model and a training method are provided. The load prediction model uses a long short-term memory network as the initial model, and introduces a Seq2seq model, a spatio-temporal neural network architecture, a dynamic time warping algorithm, and a Gaussian mixture model based on the Stackelberg game algorithm. The Seq2seq model can achieve a direct mapping from the input sequence to the output sequence without complex feature engineering. The DTW algorithm can perform dynamic alignment on time series data, further enhancing the sensitivity of the model to time series changes. The Gaussian mixture model based on the Stackelberg game algorithm can model the distribution of the load rate. Combining the Stackelberg game algorithm, it can dynamically adjust the model parameters and improve the prediction accuracy. At the same time, the load prediction model can identify and predict the operating load data and output the load rate prediction threshold interval, overcoming the problem that the point prediction in the prior art cannot effectively reflect the uncertainty in load prediction and is difficult to provide reliable information on the prediction results. Through the prediction threshold interval, the uncertainty of load prediction can be quantified, providing more comprehensive information for the operation of the power system and a more flexible decision-making basis for the power system dispatching, which helps to formulate a more robust operation strategy.

[0041] The embodiment of the present invention provides a method for fusing and processing the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm. Figure 4 The figure shows a schematic flow chart of the implementation of the method for fusing and processing the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm. The method for fusing and processing the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm includes: Step S301: Load the load rate prediction threshold interval, perform differential processing on the load rate prediction threshold interval, and obtain the load rate differential sequence of the unit's time series. In the embodiments of the present invention, differential processing can capture the dynamic change trend of the load rate, reflect the increase and decrease of the load at adjacent time points, provide richer dynamic information for subsequent analysis, and convert the absolute value of the load rate into a change amount through differential processing, reducing the absolute scale difference of the data and facilitating subsequent feature fusion and analysis.

[0042] Step S302: Determine the operating load of the generator set based on the correlation coefficient of the operating correlation parameters in the load rate difference sequence and the correlation feature set. (4) (5) Wherein, is the operating load of the generator set at time is the rated active power of the generator set, is the mean value of the load rate prediction threshold interval at time represents the mean value of the load rate prediction threshold interval at time represents the load rate difference value of the generator set at time are respectively the maximum value of the correlation coefficient of the operating correlation parameter and the correlation coefficient, represents the number of correlation parameter features. In this embodiment, it can be 5 - 20. By screening the features highly correlated with the load rate change through the correlation coefficient, redundant features are removed, and the accuracy and efficiency of the model are improved. Step S303: Use the aggregated equivalent method to perform difference compression on the generator set nodes, and with the minimum cost as the constraint, perform weighted summation on the operating load of the generator set and the correlation feature set to construct a feature fusion set. Wherein, the feature fusion set is expressed as: (6) Wherein, represents the output representation of the feature fusion set, is the number of generator sets, respectively represent the abandoned photovoltaic power component coefficient and the fuel cost coefficient of the generator set represents the number of nodes for difference compression of the generator set nodes, represents the threshold effect coefficient of the generator set respectively represent the maximum and minimum values of the active power output of the generator set.

[0043] In this embodiment, weighted summation is carried out with the minimum cost as the constraint condition, which can ensure that the operating cost is reduced on the premise of meeting the prediction accuracy. The construction of the feature fusion set can integrate the information of the operating load and the associated feature set, improving the adaptability to complex working conditions.

[0044] In the embodiment of the present invention, when constructing the optimization allocation function with the unit operating cost, the primary frequency regulation success rate, and the response time as constraints, the optimization allocation function is expressed as: (7) Wherein, represents the constraints with the unit operating cost, the primary frequency regulation success rate, and the response time, the load allocation value allocated to the generator set based on the optimization allocation function at time, are the unit operating cost, the frequency regulation dispatching cost, and the response cost respectively, are the primary frequency regulation constraint coefficient and the response time constraint coefficient respectively. In this embodiment, the primary frequency regulation constraint coefficient can be 0.85 - 0.9, and the response time constraint coefficient can be 0.8 - 1. The optimization allocation function realizes the multi-objective optimization of load allocation by comprehensively considering multiple constraint conditions such as the unit operating cost, the primary frequency regulation success rate, and the response time. It not only improves the operating efficiency and economy of the system, but also enhances the stability and dispatching flexibility of the power system. The optimization allocation function provides scientific decision-making support for power system dispatching, ensuring the efficient and stable operation of the generator set under complex working conditions. The optimization allocation function can effectively cope with uncertainty factors such as load prediction errors and equipment failures, ensuring the reliability of load allocation.

[0045] It should be noted that the primary frequency regulation constraint coefficient is a parameter used to measure the contribution degree of the generator set to the grid frequency stability during the primary frequency regulation process. It reflects the ability of the generator set to quickly adjust the active power when the grid frequency deviates from the rated value. The response time constraint coefficient is a parameter used to measure the response speed of the generator set during the primary frequency regulation process. It reflects the time required for the generator set to actually adjust the active power from detecting the frequency change. The primary frequency regulation constraint coefficient can quantify the contribution of the generator set to the grid frequency stability, ensuring that the generator set can quickly respond and adjust the active power when the grid frequency deviates from the rated value.

[0046] The embodiment of the present invention provides a load optimization control system for a generator set, Figure 5 showing a structural schematic diagram of the load optimization control system for a generator set. The load optimization control system for a generator set includes: Feature optimization module 100, which is used to obtain the operation-related parameters of the generator set on the previous day, select the operation-related parameters based on the feature priority strategy to obtain an associated feature set, construct a load forecasting model based on the long short-term memory network, and iteratively train the load forecasting model using the associated feature set; Operation data acquisition module 200, which is used to collect the operation load data of the generator set in real time, normalize the operation load data, and the load forecasting model identifies and forecasts the operation load data to output the load rate forecasting threshold interval of the generator set; Feature fusion module 300, which is used to load the load rate forecasting threshold interval and the associated feature set, and fuse and process the load rate forecasting threshold interval and the associated feature set based on the differential feature extraction algorithm to obtain a feature fusion set; Optimization control module 400, which is used to construct an optimization allocation function with the operation cost, primary frequency modulation success rate, and response time of the unit as constraints, calculate the load allocation value corresponding to the generator set during the forecasting period based on the optimization allocation function combined with the feature fusion set, trigger a load allocation instruction, and in response to the load allocation instruction, the PID control module adjusts the operation load of the generator set based on the load allocation instruction.

[0047] In this embodiment, the feature optimization module 100 includes: Parameter acquisition unit 110, which is used to obtain the operation-related parameters of the generator set on the previous day; Feature optimization unit 120, which selects the operation-related parameters based on the feature priority strategy to obtain an associated feature set; Load forecasting unit 130, which constructs a load forecasting model based on the long short-term memory network, iteratively trains the load forecasting model using the associated feature set, and outputs a converged load forecasting model.

[0048] It should be noted that the load optimization control system for the generator set provided in the embodiment of the present invention corresponds to the above-mentioned load optimization control method for the generator set. The explanations, examples, beneficial effects, etc. of the relevant content can refer to the corresponding content in the load optimization control method for the generator set, and will not be elaborated here.

[0049] In summary, the present invention provides a load optimization control method and system for a generator set. In the embodiments of the present invention, based on the feature priority strategy, the operation-related parameters are selected to obtain an association feature set that has a high degree of association with different types of generator sets and has a great influence on the load factor of the generator set, thereby ensuring the accuracy of the load prediction model for predicting the load factor of the generator set and the robustness of the model, improving the adaptability of the load prediction model to different types of generator sets, and ensuring the accuracy of adjusting the operating load of the generator set. It overcomes the problems of the existing method that it is difficult to capture complex dynamic changes by calculating and setting the load based on the difference between the theoretical load and the actual load, and has limited adaptability to different types of generator sets.

[0050] It should be noted that for the foregoing embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, some steps may be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the protection scope of the invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still, without conflict, make combinations, additions, deletions, or other adjustments to the features in the embodiments of the present invention according to the situation without creative efforts, so as to obtain different technical solutions that are essentially not divorced from the concept of the present invention, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A load optimization control method for a generator set, characterized in that, Including: Obtain the operation - related parameters of the generator set on the current day, select the operation - related parameters based on the feature - priority strategy to obtain an associated feature set, construct a load - forecasting model based on the long - short - term memory network, and iteratively train the load - forecasting model using the associated feature set; Collect the operation load data of the generator set in real - time, perform normalization processing on the operation load data, and the load - forecasting model identifies and forecasts the operation load data to output the load - rate prediction threshold interval of the generator set; Load the load - rate prediction threshold interval and the associated feature set, and perform fusion processing on the load - rate prediction threshold interval and the associated feature set based on the differential - feature extraction algorithm to obtain a feature - fusion set; Construct an optimization - allocation function with the operation cost of the unit, the success rate of primary frequency modulation, and the response time as constraints, and calculate the load - allocation value corresponding to the generator set within the prediction period based on the optimization - allocation function combined with the feature - fusion set to trigger a load - allocation instruction; In response to the load - allocation instruction, the PID control module adjusts the operation load of the generator set based on the load - allocation instruction.

2. The load optimization control method for a generator set according to claim 1, characterized in that: The method for selecting the operation - related parameters based on the feature - priority strategy includes: Load the operation - related parameters, where the operation - related parameters include unit type, active power, reactive power, load rate, load type, grid frequency, frequency fluctuation, environmental temperature, environmental humidity, wind force, altitude, rated power of the unit, power factor, fuel consumption, equipment failure rate, equipment risk, and number of shutdowns. Construct a two - dimensional feature - attribute system based on the operation - related parameters. The two - dimensional feature - attribute system consists of three types of attributes based on the time dimension and three types of attributes based on the load rate. The two - dimensional feature - attribute system divides the comprehensive parameters according to timeliness, which are the static - fact attribute matrix, the dynamic - impact attribute matrix, and the associated - interaction attribute matrix in sequence; Grab and distinguish the operation - related parameters based on the two - dimensional feature - attribute system, fill the static - fact attribute matrix, the dynamic - impact attribute matrix, and the associated - interaction attribute matrix, and assign weights to the operation - related parameters in the static - fact attribute matrix and the dynamic - impact attribute matrix respectively based on the principal - component analysis method; Load the static - fact attribute matrix and the dynamic - impact attribute matrix after weight assignment of the operation - related parameters, multiply the static - fact attribute matrix and the dynamic - impact attribute matrix by the associated - interaction attribute matrix respectively to obtain a static - association matrix and a dynamic - association matrix; Load the static - association matrix and the dynamic - association matrix, accumulate the static - association matrix and the dynamic - association matrix based on Pandas to obtain an attribute - association matrix, and calculate the correlation coefficient between the operation - related parameters and the load rate based on the Pearson algorithm; Define a parameter - determination threshold by combining the static - association matrix and the dynamic - association matrix with the comprehensive feature - priority strategy; Load the correlation coefficients of the operation - related parameters, select the operation - related parameters with a correlation coefficient greater than the parameter - determination threshold, and integrate them to obtain an associated feature set.

3. The load optimization control method for a generator set according to claim 2, wherein: The correlation coefficient between the operation - related parameters and the load rate is calculated by the following formula: （1） Among them, represents the correlation coefficient between the operation correlation parameter and the load rate, are respectively the attribute correlation matrix the load rate of the correlation parameter and the eigenvalue of the correlation parameter, is the mean value of the attribute correlation matrix, are respectively the weight values of the correlation parameter feature in the static correlation matrix and the dynamic correlation matrix; The definition calculation formula of the parameter - determination threshold is as follows: （2） （3） Among them, represents the parameter determination threshold value, are respectively the static correlation matrix and the mean value of the static correlation matrix, are respectively the dynamic correlation matrix and the mean value of the dynamic correlation matrix, respectively represent the regression coefficients of the static correlation matrix and the dynamic correlation matrix, is the minimum value of the sum of the static correlation matrix and the dynamic correlation matrix superimposed with the regression coefficient, represents the number of correlation parameter features, represents the regularization parameter of the feature priority strategy, are respectively the maximum value, minimum value, and mean value of the weights of the dynamic correlation matrix, are respectively the maximum value, minimum value, and mean value of the weights of the static correlation matrix.

4. The load optimization control method for a generator set according to claim 1, characterized in that: The method for iteratively training the load - forecasting model using the associated feature set includes: Load the associated feature set, process the associated feature set using the time-domain simulation method, and standardize the associated feature set using the max-min method. Divide the associated feature set into a training set and a test set; Load the pre-constructed load prediction model, determine the initial values of the hyperparameters of the load prediction model based on the improved vulture search-expectation maximization algorithm, and set the transfer functions of the input gate and the Seq2seq model, the transfer function between the Seq2seq model and the Gaussian mixture model, and set the loss function of the load prediction model; Obtain the training set, set the learning rate of the load prediction model to 0.002, and iteratively train the load prediction model with the training set for 100 rounds to output the initial rough model; Load the initial rough model, gradually eliminate the prediction results of the initial rough model through a multi-step inverse diffusion process, set the learning rate of the load prediction model to 0.0002, and iteratively train the load prediction model with the training set and the prediction results of the initial rough model for 50 rounds to output the initial fine model. During model training, optimize the Seq2seq model using the Adam gradient optimization algorithm, optimize the Gaussian mixture model using the variational Bayesian method, and iteratively update the hyperparameters of the load prediction model until the loss function converges to obtain the trained load prediction model; Obtain the test set, use the layer-by-layer search method to input the test set into the trained load prediction model for layer-by-layer separate testing, output the test results, and determine whether the accuracy of the test results and the true results meets the preset accuracy threshold; If the accuracy of the test results and the true results meets the preset accuracy threshold, output the converged load prediction model.

5. The load optimization control method for a generator set according to claim 4, characterized in that: The load prediction model uses a long short-term memory network as the initial model. The initial model includes an input gate, a forget gate, and an output gate connected in sequence. When constructing the load prediction model, replace the forget gate of the initial model with a Seq2seq model. The Seq2seq model includes an encoder and a decoder. Freeze the decoder of the Seq2seq model, replace the decoder with a spatio-temporal neural network architecture, and introduce the dynamic time warping algorithm into the spatio-temporal neural network architecture. The spatio-temporal neural network architecture combined with the dynamic time warping algorithm is used to normalize the operating load data and extract the associated features associated with the operating load data. Introduce a Gaussian mixture model based on the Stackelberg game algorithm between the Seq2seq model and the output gate. The Gaussian mixture model is used to take the load rate as the leader and the associated features as the follower, and combine the kernel density estimation method to determine the load rate prediction threshold interval.

6. The load optimization control method for a generator set according to claim 1, characterized in that: The method for fusing the load rate prediction threshold interval and the associated feature set based on the differential feature extraction algorithm includes: Load the load rate prediction threshold interval, perform differential processing on the load rate prediction threshold interval to obtain the load rate difference sequence of the unit's time series; Determine the operating load of the generator set based on the correlation coefficient between the load rate difference sequence and the operating correlation parameters in the associated feature set; （4） （5） Among them, is the operating load of the generating unit at time the rated active power of the generating unit, is the mean value of the load rate prediction threshold interval at time denotes the mean value of the load rate prediction threshold interval at time denotes the load rate difference value of the generating unit at time are respectively the maximum value of the correlation coefficient and the correlation coefficient of the operation-related parameters, denotes the characteristic number of the correlation parameter; Use the aggregation equivalent method to perform difference compression on the generator set nodes, and take the minimum cost as the constraint to perform weighted summation of the operating load of the generator set and the associated feature set to construct a feature fusion set; Among them, the feature fusion set is expressed as: （6） Among them, represents the output representation of the feature fusion set, is the number of generator sets, respectively represent the curtailment photovoltaic component coefficient and the fuel cost coefficient of the generator set ; represents the number of nodes for differential compression of the generator set nodes, represents the threshold effect coefficient of the generator set, respectively represent the maximum and minimum active power outputs of the generator set.

7. The load optimization control method for a generator set according to claim 6, characterized in that: When constructing the optimization allocation function with the operating cost of the unit, the success rate of primary frequency modulation, and the response time as constraints, the optimization allocation function is expressed as: （7） Among them, represents that with the operating cost of the unit, the success rate of primary frequency regulation, and the response time as constraints, the load distribution value allocated by the generating unit based on the optimization distribution function at the moment, are the operating cost of the unit, the frequency regulation dispatching cost, and the response cost respectively, are the primary frequency regulation constraint coefficient and the response time constraint coefficient respectively.

8. A load optimization control system for a generator set, which is used to implement the load optimization control method for a generator set as described in any one of claims 1-7, characterized in that: The load optimization control system for a generator set includes: A feature optimization module, configured to obtain the operating correlation parameters of the generator set on the previous day, select the operating correlation parameters based on the feature priority strategy to obtain an associated feature set, construct a load prediction model based on a long short-term memory network, and iteratively train the load prediction model using the associated feature set; An operating data acquisition module, configured to collect the operating load data of the generator set in real time, normalize the operating load data, and the load prediction model identifies and predicts the operating load data to output the predicted threshold interval of the load rate of the generator set; A feature fusion module, which loads the predicted threshold interval of the load rate and the associated feature set, and performs fusion processing on the predicted threshold interval of the load rate and the associated feature set based on the differential feature extraction algorithm to obtain a feature fusion set; An optimization control module, configured to construct an optimization allocation function with the operating cost of the unit, the success rate of primary frequency modulation, and the response time as constraints, calculate the corresponding load allocation value of the generator set within the prediction period based on the optimization allocation function in combination with the feature fusion set, trigger a load allocation instruction, and in response to the load allocation instruction, the PID control module adjusts the operating load of the generator set based on the load allocation instruction.

9. The load optimization control system for a generator set according to claim 8, wherein: The feature optimization module includes: A parameter grabbing unit, configured to obtain the operating correlation parameters of the generator set on the previous day; A feature optimization unit, which selects the operating correlation parameters based on the feature priority strategy to obtain an associated feature set; A load prediction unit, which constructs a load prediction model based on a long short-term memory network, iteratively trains the load prediction model using the associated feature set, and outputs a converged load prediction model.

Images