Intelligent speed-regulating universal generator system

Abstract

This application relates to the field of generator technology and provides an intelligent speed-regulating general-purpose generator system. Through a global perception module, it fuses the static and dynamic characteristic parameters of the prime mover with the parameter data of the general-purpose generator, and combines this with a perception state diagram to achieve a unified representation of the entire system's state. The operating condition prediction module, based on a multi-branch structure of a long short-term memory network and combined with a graph attention network, achieves hierarchical trend prediction for short-term, medium-term, and long-term conditions, ensuring that the speed regulation strategy adapts to changes in operating conditions in advance. The target optimization module constructs operating constraints based on the static characteristic parameters of the prime mover, generates elastic constraint boundaries through Bayesian inference, and dynamically adjusts them in conjunction with the trends in operating condition evolution, effectively improving operating efficiency. The execution scheduling module verifies the timing matching of instructions through dynamic time warping and adjusts the execution gain in conjunction with the dynamic characteristic parameters of the prime mover and the load change rate, significantly enhancing the generator system's adaptive operation capability under complex operating conditions.

Description

Technical Field

[0001] This application relates to the field of generator technology, and in particular to intelligent speed-regulating general-purpose generator systems. Background Technology

[0002] With the increasing demand for power supply stability in industrial production and residential electricity consumption, general-purpose generators, as core equipment for emergency power supply or off-grid power supply, have undergone an evolution in speed regulation technology from mechanical speed regulation to electronic speed regulation. Early mechanical speed regulation relied on centrifugal governors, which adjusted the fuel supply of the prime mover through mechanical feedback to achieve a rough speed stability. Later, electronic speed regulation systems introduced sensors and PID control algorithms, which adjusted the actuator action by monitoring speed deviation in real time, thus improving the speed regulation accuracy to a certain extent.

[0003] In recent years, with the increasing integration of new energy sources and complex load scenarios, generator systems have gradually evolved towards intelligence, integrating data acquisition with simple prediction algorithms to attempt to optimize speed regulation parameters using historical data. However, traditional systems treat prime mover parameters in a fragmented manner, resulting in a one-sided description of the coordinated operation of the prime mover and generator. Furthermore, they fail to consider spatial correlations between data, making it difficult to cope with complex operating conditions such as sudden load changes and environmental fluctuations, leading to unsatisfactory prediction results. Secondly, during the target optimization process, the operating constraints of the prime mover are mostly fixed values, without dynamic adjustment based on operating condition trends. This results in over- or under-constraints during load fluctuations, affecting operating efficiency and equipment safety. Finally, closed-loop execution feedback is lacking. Most systems fail to address how to achieve coordinated perception, dynamic constraint adaptation, and speed regulation execution feedback correction between the prime mover and general-purpose generator under complex operating conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides an intelligent speed-regulating general-purpose generator system, which includes: a global perception module, a working condition prediction module, a target optimization module, and an execution scheduling module;

[0005] The global perception module is used to receive the trigger signal of the target prime mover connected to the general generator, acquire and fuse the static and dynamic characteristic parameters of the target prime mover and the parameter data of the general generator in real time to form fused power generation data, and align the fused power generation data with historical power generation data in time sequence to construct a perception state diagram.

[0006] The operating condition prediction module is used to process the changing trends of fused power generation data through a long short-term memory network to deduce the evolution trend of operating conditions.

[0007] The target optimization module is used to receive the static characteristic parameters of the target prime mover to determine the operating constraints of the target prime mover, analyze the user-preset optimization target, and generate an instruction sequence including speed trajectory and actuator action based on the perception state diagram and the working condition evolution trend through reinforcement learning.

[0008] The execution scheduling module is used to receive and verify the instruction sequence, and dynamically adjust the execution gain of the instruction sequence according to the load change rate of the general generator and the dynamic characteristic parameters of the target prime mover, and monitor the execution status of the instruction sequence to update the perception status map.

[0009] As an optional implementation, the construction logic of the perception state diagram includes:

[0010] The merged power generation data is time-series aligned with historical power generation data, and non-uniform time series interpolation is performed on the merged power generation data through dynamic time warping.

[0011] The fused power generation data is divided by adaptive time windows to extract local state features. The local state features within each time window constitute the nodes of the perception state map.

[0012] The cosine similarity and dynamic time bending distance of adjacent nodes are calculated to determine the edges and connection weights of adjacent nodes. Oriented edges of non-adjacent nodes are added through a causal inference algorithm to construct a perceptual state graph.

[0013] As an optional implementation, the sub-logic for forming the fused power generation data includes:

[0014] The static characteristic parameters of the target prime mover are characterized and anchored, and the dynamic characteristic parameters of the target prime mover and the parameter data of the general generator are subjected to multi-mode filtering.

[0015] The data type is determined and the correlation of data of the same type is aggregated. At the same time, the data of different types are fused across domains through a heterogeneous graph neural network. The static characteristic parameters of the target prime mover are used as the initial node features of the heterogeneous graph neural network.

[0016] The dynamic characteristic parameters of the target prime mover and the parameter data of the general generator are used as edge features, and the interaction weights of different types of data are learned through an attention mechanism to form fused power generation data.

[0017] As an optional implementation, the deduction logic for the evolution trend of the operating conditions includes:

[0018] The integrated power generation data is divided into high-frequency and low-frequency components according to the time dimension, and the correlation between the integrated power generation data is sorted out according to the spatial dimension. The integrated power generation data is then calculated by mutual information entropy to filter out strongly correlated data pairs.

[0019] Long Short-Term Memory (LSTM) networks consist of short-term, medium-term, and long-term branches. High-frequency components are input to the short-term branch to output short-term prediction sequences, low-frequency components and short-term prediction sequences are input to the medium-term branch to output phase prediction sequences, and strongly correlated data pairs and phase prediction sequences are input to the long-term branch to output trend prediction sequences.

[0020] The branch weights are dynamically adjusted according to the complexity of the working conditions. After fusion, a preliminary trend sequence is formed. The perception state map is received and the correlation strength between nodes is calculated through the graph attention network to generate a topology vector. The preliminary trend sequence and the topology vector are then fused to generate a trend matrix.

[0021] By processing the trend matrix and historical prediction errors through Bayesian inference, the confidence interval of each prediction result is calculated. At the same time, the confidence interval of the prediction result is corrected according to the degree centrality of the nodes in the perception state graph to infer the evolution trend of the working condition. The difference between the evolution trend of the working condition and the actual data is used to determine whether to adjust the branch weights of the long short-term memory network.

[0022] As an optional implementation, the logic for generating the instruction sequence includes:

[0023] The local state features and edge connection weights in the perception state graph are used to generate an initial state vector, and then the working condition evolution trend weights are assigned through an attention mechanism and fused into a state vector.

[0024] The algorithm analyzes user-preset optimization goals and generates basic rewards, and applies negative rewards to actions that violate operational constraints in order to determine the reward function for reinforcement learning.

[0025] Based on the state vector and reward function, candidate speed trajectories are generated through Monte Carlo tree search. The candidate speed trajectories are verified by running constraints to determine the speed trajectory. Based on the speed trajectory, the executor actions are generated through the near-end policy algorithm to generate the instruction sequence.

[0026] As an optional implementation, the sub-logic for determining the operational constraints includes:

[0027] Receive the static characteristic parameters of the target prime mover, set a multi-dimensional constraint space based on the rated power, efficiency curve and structural parameters, and generate initial constraint conditions in combination with the type of the target prime mover;

[0028] Based on historical operating data, elastic constraint boundaries are set for each initial constraint condition, and the confidence level of the elastic constraint boundaries is calculated through Bayesian inference.

[0029] The elastic constraint boundary is dynamically adjusted according to the evolution trend of the operating conditions in order to determine the operating constraints of the target prime mover.

[0030] As an optional implementation, the feedback update logic of the perception state diagram includes:

[0031] The parameter data of the general generator during the execution of the instruction sequence is obtained, and the parameter data is compared with the local state features in the perception state diagram to calculate the feature deviation.

[0032] Configure a deviation threshold. When the feature deviation is less than the deviation threshold, replace the node with the local state feature in the perception state graph, and update the connection weight between the nodes according to the execution state of the instruction sequence.

[0033] When the characteristic deviation is greater than or equal to the deviation threshold, the cause of the deviation is analyzed by combining the actuator action and the load change rate of the general generator, so as to determine whether to add a new directional edge in the perception state diagram.

[0034] The updated node and edge connection weights are time-series aligned with historical power generation data to verify the rationality of the perception state graph update.

[0035] As an optional implementation, the verification logic of the instruction sequence includes:

[0036] Receive instruction sequences, extract timestamps of actuator actions and time nodes of rotational speed trajectories, and verify the timing matching degree between actuator actions and rotational speed trajectories through dynamic time warping;

[0037] Call the elastic constraint boundary in the running constraints, and compare the actuator action and speed trajectory with the elastic constraint boundary to determine the cause of the over-limit;

[0038] By combining the dynamic characteristic parameters of the target prime mover, the execution of the command sequence is simulated, and the mechanical state of the general-purpose generator is monitored during the simulation to determine whether the command sequence needs to be corrected.

[0039] As an optional implementation, the gain adjustment sub-logic includes:

[0040] The load change rate of the general generator is acquired in real time, and the correlation analysis between the load change rate and the speed fluctuation of the target prime mover is performed to generate a load-speed correlation map.

[0041] Based on the load-speed correlation graph, the execution gain of the command sequence is determined by combining the torque response delay of the target prime mover, and the command sequence is adjusted based on the execution gain;

[0042] The execution status of the instruction sequence is continuously monitored, and the difference between the actual execution effect and the expected execution effect is calculated to determine whether the execution gain of the instruction sequence should be adjusted again in combination with the vibration frequency.

[0043] As an optional implementation, the static characteristic parameters of the target prime mover include rated power, efficiency curve, and structural parameters; the dynamic characteristic parameters of the target prime mover include speed fluctuation, torque response delay, and vibration frequency; the parameter data of the general-purpose generator includes electrical parameters, mechanical status, environmental information, and load characteristics; the electrical parameters of the general-purpose generator include generator speed, voltage harmonic distortion rate, and power factor; the mechanical status of the general-purpose generator includes bearing temperature and vibration spectrum; the environmental information of the general-purpose generator includes ambient temperature and ambient humidity; and the load characteristics of the general-purpose generator include load change rate and power distribution timing.

[0044] Compared with existing technologies, the beneficial effects of this application are as follows: By integrating the static and dynamic characteristic parameters of the prime mover and the parameter data of the general generator through the global perception module, and combining them with the construction of the perception state diagram, a unified representation of the entire system state is achieved, which greatly improves the completeness of the state description and provides comprehensive data support for subsequent optimization decisions; The operating condition prediction module is based on the multi-branch structure of the long short-term memory network and combines it with the graph attention network to achieve hierarchical prediction of short-term instantaneous fluctuations, medium-term trend changes and long-term evolution laws, and corrects the confidence interval through Bayesian inference, which significantly improves the prediction accuracy and ensures that the speed regulation strategy adapts to the evolution of operating conditions in advance.

[0045] The target optimization module constructs a multi-dimensional constraint space based on the static characteristic parameters of the prime mover, generates elastic constraint boundaries through Bayesian inference, and dynamically adjusts them in conjunction with the evolution trend of the operating conditions. During the reinforcement learning process, the reward function incorporates both the optimization objective and the constraint violation penalty. The generated instruction sequence effectively improves operating efficiency while ensuring equipment safety, resolving the contradiction between efficiency and safety under fixed constraints. The execution scheduling module verifies the timing matching of instructions through dynamic time warping, adjusts the execution gain in conjunction with the dynamic characteristic parameters of the prime mover and the load change rate, and establishes a real-time update mechanism for the perception state diagram. This accelerates the generator system's response speed to load changes, reduces equipment losses caused by abnormal vibration frequencies, and significantly enhances the generator system's adaptive operation capability under complex operating conditions. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0047] Figure 1 This is a system flowchart of the intelligent speed-regulating universal generator system provided in the embodiments of this application;

[0048] Figure 2A logical diagram illustrating the evolution of operating conditions of the intelligent speed-regulating universal generator system provided in this application embodiment;

[0049] Figure 3 The sub-logic diagram for determining the operating constraints of the intelligent speed-regulating general-purpose generator system provided in the embodiments of this application is shown below.

[0050] Figure 4 This is a feedback update logic diagram of the perception state diagram of the intelligent speed-regulating universal generator system provided in the embodiments of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0052] Example

[0053] like Figure 1 The diagram shown is a system flowchart of an intelligent speed-regulating general-purpose generator system provided in this application embodiment. The system includes a global perception module, a working condition prediction module, a target optimization module, and an execution scheduling module.

[0054] The global perception module is used to receive the trigger signal when the target prime mover is connected to the general generator, acquire and fuse the static and dynamic characteristic parameters of the target prime mover and the parameter data of the general generator in real time to form fused power generation data, align the fused power generation data with historical power generation data in time sequence and construct a perception state diagram.

[0055] The static characteristic parameters of the target prime mover include rated power, efficiency curve, and structural parameters. The dynamic characteristic parameters of the target prime mover include speed fluctuation, torque response delay, and vibration frequency. The parameter data of the general generator include electrical parameters, mechanical condition, environmental information, and load characteristics. The electrical parameters of the general generator include generator speed, voltage harmonic distortion rate, and power factor. The mechanical condition of the general generator includes bearing temperature and vibration spectrum. The environmental information of the general generator includes ambient temperature and ambient humidity. The load characteristics of the general generator include load change rate and power matching sequence.

[0056] It needs to be explained that: rated power refers to the maximum power that the prime mover can continuously output under standard design conditions. It is the basic indicator for measuring the prime mover's work capacity. Standard operating conditions refer to standard ambient temperature and rated speed. The efficiency curve reflects the efficiency change of the prime mover under different load rates, usually showing a trend of high efficiency in the middle and low efficiency at both ends. The load rate refers to the ratio of actual output power to rated power. Structural parameters cover the mechanical construction characteristics of the prime mover, including the number of cylinders, stroke length, rotor diameter, and bearing type, which directly affect the mechanical strength, heat dissipation capacity, and operational stability of the prime mover. Speed ​​fluctuation refers to the fluctuation range of the actual speed of the prime mover from the rated speed when the load changes. The smaller the fluctuation, the stronger the adaptability of the prime mover to load changes. Torque response delay refers to the time difference between the current value and the target value of the output torque of the prime mover when the load demand changes. The smaller the delay, the more sensitive the dynamic response of the prime mover. Vibration frequency refers to the frequency components of the vibration signal generated by the mechanical motion of the prime mover during operation. Different faults will lead to an increase in vibration at a specific frequency.

[0057] Generator speed refers to the rotational speed of the generator rotor, which directly affects the frequency of the output voltage; voltage harmonic distortion rate refers to the ratio of harmonic components to fundamental components in the output voltage, reflecting the degree of voltage waveform distortion. Excessive distortion rate can affect the lifespan of electrical equipment; power factor represents the ratio of active power to apparent power, measuring energy utilization efficiency. A low factor increases line losses and generator load; bearing temperature refers to the operating temperature of the generator bearings, a key indicator for judging bearing wear and lubrication status. Exceeding the threshold can lead to bearing burnout; vibration spectrum represents the frequency distribution of vibration signals during generator operation, including operating status information of components such as the rotor and stator, and can be used for early fault diagnosis; ambient temperature refers to the air temperature of the generator's environment, affecting heat dissipation efficiency and equipment insulation performance; ambient humidity refers to the water vapor content in the ambient air. Excessive humidity can cause electrical components to become damp and short-circuit, while insufficient humidity can cause electrostatic interference; load change rate represents the magnitude of load power change per unit time, reflecting load volatility, i.e., a large rate of change during sudden load changes and a small rate of change during stable load changes; power distribution sequence refers to the changing pattern of the power proportion of different types of loads over time, affecting the generator's reactive power demand.

[0058] Furthermore, the sub-logic for forming integrated power generation data includes:

[0059] The static characteristic parameters of the target prime mover are characterized and anchored, and the dynamic characteristic parameters of the target prime mover and the parameter data of the general generator are subjected to multi-mode filtering.

[0060] The data type is determined and the correlation of data of the same type is aggregated. At the same time, the data of different types are fused across domains through a heterogeneous graph neural network. The static characteristic parameters of the target prime mover are used as the initial node features of the heterogeneous graph neural network.

[0061] The dynamic characteristic parameters of the target prime mover and the parameter data of the general generator are used as edge features, and the interaction weights of different types of data are learned through an attention mechanism to form fused power generation data.

[0062] The static characteristic parameters of the target prime mover come from different models of equipment, and the parameter formats and dimensions differ. Direct use of these parameters can affect the accuracy of data fusion. Dynamic characteristic parameters and parameters from general-purpose generators are susceptible to noise interference during acquisition and require filtering to ensure data quality. For static characteristic parameters, a standardized parameter mapping library is constructed. Parameters such as rated power and efficiency curves of different prime mover models are converted into feature vectors of a unified dimension according to preset rules. For example, discrete points of the efficiency curve are converted into fixed-length vectors through interpolation to complete feature anchoring. For dynamic characteristic parameters and parameters from general-purpose generators, multimodal filtering is used. High-frequency signals such as vibration frequency are denoised using wavelet filtering, electrical parameters such as voltage harmonic distortion rate are filtered using Kalman filtering to eliminate measurement noise, and slowly changing parameters such as ambient temperature and humidity are filtered using moving average filtering. Through feature anchoring, the format differences of static parameters from different prime mover models are eliminated, providing a unified data foundation for subsequent cross-domain fusion. Multimodal filtering addresses the characteristics of different data types, effectively reducing noise interference and improving data reliability.

[0063] The data types are diverse, and there are inherent correlations among data of the same type. Aggregation can enhance feature expression. There are also mutual influences between different types of data, such as environmental temperature affecting mechanical state. Cross-domain fusion can uncover potential correlations and improve the comprehensiveness of the data. The filtered dynamic characteristic parameters and parameters of the general generator are divided into mechanical, electrical, environmental, and load types. For data of the same type, the correlation degree between data is calculated through correlation analysis algorithms. Data with high correlation degree are aggregated. For example, bearing temperature and vibration spectrum in mechanical state are aggregated into mechanical health feature vector. For cross-domain fusion, a heterogeneous graph neural network is constructed. The feature vectors corresponding to static characteristic parameters are used as initial node features, and the dynamic characteristic parameters of different types of target prime movers and parameters of the general generator are used as edge features input to the heterogeneous graph neural network. The interaction relationship between different types of data is learned through inter-layer propagation of the heterogeneous graph neural network. The correlation aggregation of data of the same type reduces data dimensionality and highlights key features. The cross-domain fusion of heterogeneous graph neural network fully explores the potential correlations between different types of data, forming more comprehensive and representative data features, providing rich information for the subsequent construction of the perception state map.

[0064] Different types of data have varying degrees of impact on the generator system state. For example, changes in load characteristics have a greater impact on the generator system than changes in environmental information, requiring different weights to highlight the role of key data. In heterogeneous graph neural networks, dynamic characteristic parameters and various parameters of the general generator are used as edge features. An attention mechanism is used to analyze the contribution of different types of data to the changes in the general generator system state during historical operation, automatically learning and assigning interaction weights. For example, when the load change rate is large, the weight of load features in the edge features is increased. Then, the weighted edge features are fused with the initial node features to form the final fused power generation data. The attention mechanism can adaptively assign weights to different types of data, making the fused power generation data more reflective of the key state changes of the system and improving the information value of the data. As the foundation for constructing the perception state diagram, the fused power generation data directly affects the accuracy of the state diagram. High-quality fused data enables the subsequent perception state diagram to more accurately reflect the operating state of the generator system.

[0065] Specifically, the construction logic of the perception state diagram includes:

[0066] The merged power generation data is time-series aligned with historical power generation data, and non-uniform time series interpolation is performed on the merged power generation data through dynamic time warping.

[0067] The fused power generation data is divided by adaptive time windows to extract local state features. The local state features within each time window constitute the nodes of the perception state map.

[0068] The cosine similarity and dynamic time bending distance of adjacent nodes are calculated to determine the edges and connection weights of adjacent nodes. Oriented edges of non-adjacent nodes are added through a causal inference algorithm to construct a perceptual state graph.

[0069] The different acquisition time intervals of the merged power generation data and historical power generation data lead to temporal misalignment, affecting data comparison and analysis. Furthermore, the merged power generation data exhibits uneven sampling, requiring interpolation to ensure data continuity. Using the generator system's operating timeline as a benchmark, the timestamps of the merged power generation data and historical power generation data are matched. For data with mismatched time series, a dynamic time warping algorithm is used to adjust the time series of the merged power generation data, achieving temporal alignment. For non-uniform sampling points in the merged power generation data, dynamic time warping is used for interpolation, generating interpolated data at intermediate times based on the changing trends of adjacent sampling points, ensuring a uniform distribution of the merged power generation data across the time dimension. Temporal alignment resolves the temporal misalignment issue from different data sources, facilitating comparative analysis between historical and current merged power generation data. Non-uniform time series interpolation ensures data continuity, providing a coherent data foundation for subsequent local state feature extraction.

[0070] The operating state of a generator system changes over time, with different state characteristics in different time segments. Fixed time window divisions cannot accurately capture key state changes. Adaptive time window division adjusts the window length based on the intensity of state changes. When the rate of change in the fused power generation data exceeds a preset threshold, the time window length is automatically shortened to capture rapidly changing states more precisely. Conversely, when the rate of change is less than the preset threshold, the time window length is extended to reduce data processing volume. Within each time window, local state features such as the mean, trend, and characteristic peaks are extracted, and each local state feature serves as a node in the perception state map. Adaptive time window division flexibly adjusts the window length according to changes in the generator system state, ensuring precise capture of rapidly changing states while improving data processing efficiency under stable conditions. The extracted local state features serve as nodes, providing basic building blocks for the perception state map.

[0071] Local state features in adjacent time windows exhibit certain correlations, which can be quantified using similarity and distance calculations and used as connection weights for edges. However, indirect causal relationships exist between non-adjacent nodes, requiring causal inference mining to refine the structure of the perception state graph. Cosine similarity between adjacent nodes is calculated to measure the similarity between local state features, while dynamic time curvature distance is calculated to measure the matching degree of time series. These two factors are combined as connection weights for adjacent node edges. For non-adjacent nodes, causal inference algorithms analyze the causal relationships of data changes within the time windows corresponding to different nodes. When significant causal correlations exist, directional edges are added, and weights are assigned based on the degree of causal influence. The connection weights of adjacent node edges accurately reflect the correlation degree of temporally continuous local state features, while directional edges of non-adjacent nodes mine potential causal relationships. This ensures that the perception state graph includes not only temporally continuous correlations but also non-continuous causal correlations, providing a more comprehensive reflection of the relationships between generator system states.

[0072] A structured approach is needed to integrate the local state features and their relationships within a generator system, providing a clear and comprehensive view of its overall operating status. Extracted local state features are used as nodes, and edges and connection weights of adjacent nodes, as well as directional edges of non-adjacent nodes, are organized according to their relationships to form a complete perceptual state graph. This graph includes node and edge attribute information. Node attributes include local state features, while edge attributes include connection weights and causal relationships. The perceptual state graph graphically integrates the generator system's state features and relationships, making the complex operating status of the generator system intuitive and easy to understand, facilitating analysis and processing by subsequent modules. The perceptual state graph is a crucial basis for the operating condition prediction module to process the changing trends of fused generator data and infer the evolution of operating conditions. A high-quality perceptual state graph can significantly improve the accuracy of operating condition predictions.

[0073] The operating condition prediction module is used to process the changing trends of fused power generation data through a long short-term memory network to deduce the evolution trend of operating conditions.

[0074] Specifically, such as Figure 2 As shown, the deductive logic for the evolution trend of operating conditions includes:

[0075] The integrated power generation data is divided into high-frequency and low-frequency components according to the time dimension, and the correlation between the integrated power generation data is sorted out according to the spatial dimension. The integrated power generation data is then calculated by mutual information entropy to filter out strongly correlated data pairs.

[0076] Long Short-Term Memory (LSTM) networks consist of short-term, medium-term, and long-term branches. High-frequency components are input to the short-term branch to output short-term prediction sequences, low-frequency components and short-term prediction sequences are input to the medium-term branch to output phase prediction sequences, and strongly correlated data pairs and phase prediction sequences are input to the long-term branch to output trend prediction sequences.

[0077] The branch weights are dynamically adjusted according to the complexity of the working conditions. After fusion, a preliminary trend sequence is formed. The perception state map is received and the correlation strength between nodes is calculated through the graph attention network to generate a topology vector. The preliminary trend sequence and the topology vector are then fused to generate a trend matrix.

[0078] By processing the trend matrix and historical prediction errors through Bayesian inference, the confidence interval of each prediction result is calculated. At the same time, the confidence interval of the prediction result is corrected according to the degree centrality of the nodes in the perception state graph to infer the evolution trend of the working condition. The difference between the evolution trend of the working condition and the actual data is used to determine whether to adjust the branch weights of the long short-term memory network.

[0079] The integrated power generation data contains various data points, some changing rapidly and others slowly. Processing them together can mask the trends of slower changes with the fluctuations of faster ones. Furthermore, some data points have close relationships while others have loose ones; therefore, closely related parameter pairs must be selected to avoid interference from irrelevant information. In the time dimension, wavelet transform is used to divide the integrated power generation data into two parts: high-frequency components correspond to rapidly changing data, including load change rate, preserving their abrupt changes; low-frequency components correspond to slowly changing parameters, including ambient temperature and bearing temperature, extracting their long-term trend characteristics. In the spatial dimension, mutual information entropy is used to calculate the correlation between any two data points, such as determining the strength of the correlation between speed fluctuations and torque response delay, and between load change rate and power factor. The strongest correlations are selected, filtering out strongly correlated data pairs, such as load change rate and generator speed, and bearing temperature and vibration spectrum, and labeling them as causal or consequential relationships. Temporal splitting preserves the characteristics of fast and slow changes independently, preventing mutual interference. Spatial filtering identifies closely related data pairs, reducing the influence of irrelevant parameters, allowing the Long Short-Term Memory network to focus on parameter relationships that play a key role in changes in operating conditions, making the input data more targeted.

[0080] The causes of operating condition changes vary across different time periods. In the short term, they are mainly affected by sudden load changes and instantaneous interference; in the medium term, they are affected by environmental changes and accumulated heat from equipment; and in the long term, they are affected by the cumulative effects of equipment aging and wear. Therefore, different branches need to be processed, and the importance of each branch varies under different operating conditions. During sudden loading, the short-term branch is more important, while during stable operation, the long-term branch is more important. Thus, the weights of each branch need to be adjusted. The short-term branch receives high-frequency components and uses a network structure with a forget gate. When the load changes drastically, the weight of the forget gate is reduced, allowing the Long Short-Term Memory (LSTM) network to focus more on the latest high-frequency data, outputting the instantaneous operating condition prediction results for a short period, resulting in a short-term prediction sequence, including rotational speed. The peak value and the timing of sudden power changes are analyzed. The intermediate branch receives the low-frequency components and the output of the short-term branch. Using a bidirectional network structure, it learns the historical slow trends forward and predicts the future direction of change backward. Combining the changes in ambient temperature and the cumulative bearing temperature, it outputs the stage condition prediction results within the intermediate time period, resulting in a stage prediction sequence, including the trend of temperature changes and the gradual change in efficiency. The long-term branch receives closely related parameter pairs and the output of the intermediate branch. Referring to similar historical cases of operating condition changes, it corrects the prediction results using the change paths of these cases, outputting the long-term trend prediction results over a long period, resulting in a trend prediction sequence, including the nodes where the equipment shows aging warnings and the time when efficiency drops to the critical value.

[0081] Dynamic weight adjustment is performed by an evaluator of operational complexity. This evaluator adjusts the weights based on the severity of data fluctuations, which considers both the proportion of rapidly changing data and the frequency of sudden changes. When operational conditions are complex—i.e., sudden load changes or simultaneous fluctuations in multiple parameters—the weight of short-term branches is increased. When operational conditions are stable, the weight of long-term branches is increased. After adjustment, the prediction results from each branch are weighted and merged to form a preliminary trend sequence, along with a reliability score for each branch. This multi-branch structure can learn the patterns of operational condition changes over different time periods, resulting in more accurate predictions. Dynamic weight adjustment allows the merged preliminary trend sequence to adapt to varying operational complexities, focusing on instantaneous changes during complex conditions and on long-term trends during stable conditions, thus enhancing its adaptability.

[0082] The preliminary trend sequence only reflects the changes in data over time, without showing the spatial correlation and overall structure between data. However, the node features and edge connection weights in the perception state graph contain the spatial driving relationship of the changes in operating conditions. After fusion, the prediction results can be more interpretable. After receiving the perception state graph, a graph attention network is used to calculate the correlation strength between nodes. Core nodes with more connections to other nodes are given higher attention weights, while edge nodes with fewer connections are given lower weights. Core nodes include generator speed and load change rate, while edge nodes include ambient humidity. The weighted result of node features and attention weights is extracted as node importance features. At the same time, edge connection weights and causal labels are extracted to generate a correlation strength matrix. The node importance features and correlation strength matrix are integrated to form a topology vector.

[0083] The preliminary trend sequence and topology vector are fused through a gating fusion mechanism. When there is a strong causal relationship in the perception state diagram, the proportion of the topology vector in the fusion is increased. When the predictive reliability of the generator speed in the preliminary trend sequence is high, the proportion of the preliminary trend sequence is increased. The trend matrix generated after fusion includes time dimension, data dimension, and correlation dimension. The time dimension includes short time, medium time, and long time. The parameter dimension includes electrical parameters and mechanical state, etc. The correlation dimension includes the causal strength between data. After adding the topology vector, the trend matrix can not only predict the trend of data change, but also explain the reasons for the data change, making the prediction results easier to understand. The gating fusion mechanism adjusts the fusion ratio according to the reliability of the time series prediction and the strength of spatial correlation to avoid interference from weak correlation information.

[0084] The prediction results of the trend matrix are uncertain. In unfamiliar working conditions, the prediction error is large, while in common working conditions and those dominated by core nodes, the prediction is more reliable. Bayesian inference is needed to quantify this uncertainty, and the confidence interval is adjusted according to the degree centrality of the nodes. At the same time, the deviation between the prediction results and the actual data needs to be fed back to the Long Short-Term Memory network to continuously optimize the learning methods of each branch. In Bayesian inference, the confidence interval of each prediction result in the trend matrix is ​​first calculated based on the distribution of historical prediction errors. Then, it is corrected according to the number of connections of nodes in the perception state graph. The confidence interval of core nodes with many connections is narrowed, and the confidence interval of peripheral nodes with few connections is widened. The corrected confidence interval and the prediction result together constitute the working condition evolution. The system adjusts the trend of the operating conditions during feedback. It compares the deviation between the trend of the operating conditions and the actual data. When the deviation is large, if the deviation is large in the short-term branches, the importance of high-frequency components in training is increased. If the deviation is large in the topological vector fusion, the weights of the graph attention network computing nodes are adjusted. After adjustment, samples with large deviations are marked as abnormal operating condition samples and stored in the database for subsequent training of the long short-term memory network. Bayesian inference combined with the degree centrality of the nodes corrects the confidence interval of the prediction results, making the prediction reliability of core nodes more accurate and the uncertainty of edge nodes can be reasonably reflected. The feedback adjustment mechanism gradually reduces the prediction deviation of the long short-term memory network under complex operating conditions by optimizing the training method in a targeted manner, making the long short-term memory network more stable.

[0085] The target optimization module receives the static characteristic parameters of the target prime mover to determine the operating constraints of the target prime mover, analyzes the user-preset optimization target, and generates a sequence of instructions including speed trajectory and actuator action based on the perception state diagram and the evolution trend of the operating condition through reinforcement learning.

[0086] Furthermore, such as Figure 3 As shown, the sub-logic for determining the execution constraints includes:

[0087] Receive the static characteristic parameters of the target prime mover, set a multi-dimensional constraint space based on the rated power, efficiency curve and structural parameters, and generate initial constraint conditions in combination with the type of the target prime mover;

[0088] Based on historical operating data, elastic constraint boundaries are set for each initial constraint condition, and the confidence level of the elastic constraint boundaries is calculated through Bayesian inference.

[0089] The elastic constraint boundary is dynamically adjusted according to the evolution trend of the operating conditions in order to determine the operating constraints of the target prime mover.

[0090] The static characteristic parameters of the target prime mover are its inherent attributes, determining the basic boundaries of operation. Different types of prime movers have different working principles, and the constraints need to be set accordingly. Failure to consider the type of prime mover will result in poor applicability of the constraints. After receiving the rated power, it is converted into the upper and lower limit basic values ​​of the output power. The efficiency curve is divided into high-efficiency and low-efficiency ranges. The low-efficiency range serves as a constraint no-go zone. Information such as material and size in the structural parameters is mapped to the basic thresholds of mechanical stress and heat dissipation capacity. The preset rule library is called in combination with the type of prime mover. For example, an additional exhaust temperature constraint is added for diesel prime movers, while combustion efficiency constraints are strengthened for gas prime movers. These are integrated to form the initial constraint conditions. Setting constraints based on the inherent attributes of the prime mover ensures the basicity and rationality of the operating constraints. Combining the differentiated processing of the prime mover type makes the initial operating constraints more in line with actual operating needs and provides a benchmark for setting elastic constraint boundaries. Its accuracy directly affects the accuracy of subsequent boundary adjustments and lays the foundation for the dynamic optimization of operating constraints.

[0091] Initial constraints are fixed values, but reasonable fluctuations exist in actual operation. Slight load fluctuations leading to short-term power overruns necessitate the setting of elastic constraint boundaries to accommodate normal fluctuations. Simultaneously, the reliability of these elastic constraint boundaries needs to be quantified to distinguish between reasonable fluctuations and abnormal overruns. The actual fluctuation range of each initial constraint is extracted from historical operating data; for example, the fluctuation range of the actual operating value of rated power around the base value. This range is set as the elastic constraint boundary. Through Bayesian inference, combined with the number of normal operations within the boundary and the number of abnormal operations outside the boundary in historical operating data, the confidence level of the elastic constraint boundary is calculated. For example, if a boundary can effectively distinguish between normal and abnormal operations 95% of the time in history, the confidence level is 95%. The elastic constraint boundary provides a buffer for normal operating fluctuations, avoiding decreased operating efficiency due to excessive constraints. The confidence level calculation quantifies the reliability of the elastic constraint boundary, providing a risk reference for subsequent adjustments.

[0092] The operating condition evolution trend reflects the future direction of data change. Flexible constraint boundaries need to be adjusted in advance according to this trend to reserve a safety margin; if they remain fixed, they will lose their constraint effect during trend changes. After receiving the operating condition evolution trend, the impact of this trend on each flexible constraint boundary is analyzed. For example, if the trend indicates that the ambient temperature will continue to rise, the heat dissipation-related operating constraint boundaries need to be tightened. The flexible constraint boundaries are adjusted based on the strength of the trend; a strong trend results in a larger adjustment, while a weak trend results in a smaller adjustment. The adjusted boundaries need to have their confidence levels updated again through Bayesian inference. For example, after adjusting the flexible operating boundaries driven by the operating condition evolution trend, the confidence level is recalculated based on operating results under similar historical trends to determine the operating constraints of the target prime mover. This ensures that the flexible constraint boundaries are forward-looking, proactively adapting to changing operating conditions and avoiding equipment risks caused by constraint lag during trend changes. The updated confidence levels ensure the reliability of the adjusted flexible constraint boundaries. The dynamically adjusted flexible constraint boundaries constitute the final operating constraints, providing real-time and accurate constraint basis for the generation of instruction sequences, ensuring that the instruction sequences are optimized within the safety boundaries.

[0093] Specifically, the logic for generating the instruction sequence includes:

[0094] The local state features and edge connection weights in the perception state graph are used to generate an initial state vector, and then the working condition evolution trend weights are assigned through an attention mechanism and fused into a state vector.

[0095] The algorithm analyzes user-preset optimization goals and generates basic rewards, and applies negative rewards to actions that violate operational constraints in order to determine the reward function for reinforcement learning.

[0096] Based on the state vector and reward function, candidate speed trajectories are generated through Monte Carlo tree search. The candidate speed trajectories are verified by running constraints to determine the speed trajectory. Based on the speed trajectory, the executor actions are generated through the near-end policy algorithm to generate the instruction sequence.

[0097] The perception state diagram includes the current local state and relationships of the generator system, while the operating condition evolution trend reflects the future direction of change. The fusion of these two aspects can comprehensively describe the current state and future trend of the generator system, providing complete input information for reinforcement learning. Using only a single piece of information will lead to one-sided learning. The local state features of each node in the perception state diagram are extracted, and the comprehensive influence of the nodes is calculated by combining the connection weights of the edges. This is then integrated to form an initial state vector. The importance of each data point in the operating condition evolution trend is analyzed through an attention mechanism. For example, the load change rate has a higher weight than the ambient temperature in the operating condition evolution trend. The weighted operating condition evolution trend features are fused with the initial state vector to form a state vector that includes spatiotemporal information. The fused state vector simultaneously covers the details of the current state and the guidance of the future trend, enabling reinforcement learning to take into account both current optimization and future adaptation. Its completeness directly affects the calculation of the reward function and the optimization of the policy, providing an accurate state description for the generation of instruction sequences.

[0098] Reinforcement learning requires explicit reward signals to guide strategy optimization. The reward function must reflect the user's optimization goals while penalizing behaviors that violate constraints, thus balancing optimization and safety. The process involves analyzing the user's preset optimization goals. If the goal is to maximize efficiency, the real-time efficiency value is converted into a basic reward; higher efficiency results in a larger reward. Simultaneously, actions that violate operational constraints are assigned negative rewards based on the degree of violation. For example, minor violations result in small negative rewards, while severe violations result in large negative rewards, and the absolute value of the negative reward is greater than a positive reward of the same degree, ensuring safety is prioritized. Finally, the basic and negative rewards are integrated to form the reward function. The reward function transforms abstract optimization goals and constraints into quantifiable signals, balancing optimization performance and operational safety. The reward function serves as the basis for updating the reinforcement learning strategy; its rationality determines whether the generated instruction sequence meets user needs and safety requirements, providing clear guidance for instruction sequence optimization.

[0099] The rotational speed trajectory is the core objective of the prime mover's operation and must be generated within the constraints of the operating conditions. Actuator actions are the means to achieve this trajectory and must be precisely matched with it. Only by combining the two can an executable instruction sequence be formed. Optimizing a single aspect will lead to poor overall performance. Based on state vectors and reward functions, Monte Carlo tree search is used to simulate different speed change paths, generating multiple candidate rotational speed trajectories. Operating constraints are then invoked to verify these candidate trajectories, eliminating those that exceed the constraints. Feasible trajectories are retained, and the one with the highest reward is selected as the final rotational speed trajectory. Using this trajectory as the target, a near-end strategy algorithm generates actuator actions. The timing and magnitude of the actuator actions must match the key nodes of the rotational speed trajectory; for example, the node where the rotational speed increases corresponds to the start time of the actuator action. The generation and verification of candidate rotational speed trajectories ensure the feasibility and optimizability of the rotational speed trajectory. The matching of actuator actions with the rotational speed trajectory ensures the executability of the instruction sequence, enabling the optimization objective to be implemented through specific actions. The generated instruction sequence is directly input into the execution scheduling module, and its accuracy and optimization determine the effectiveness of the execution scheduling, providing specific operational instructions for the intelligent speed regulation of the generator system.

[0100] The execution scheduling module is used to receive and verify the instruction sequence, and dynamically adjust the execution gain of the instruction sequence according to the load change rate of the general generator and the dynamic characteristic parameters of the target prime mover, and monitor the execution status of the instruction sequence to update the perception status map.

[0101] Furthermore, the verification logic of the instruction sequence includes:

[0102] Receive instruction sequences, extract timestamps of actuator actions and time nodes of rotational speed trajectories, and verify the timing matching degree between actuator actions and rotational speed trajectories through dynamic time warping;

[0103] Call the elastic constraint boundary in the running constraints, and compare the actuator action and speed trajectory with the elastic constraint boundary to determine the cause of the over-limit;

[0104] By combining the dynamic characteristic parameters of the target prime mover, the execution of the command sequence is simulated, and the mechanical state of the general-purpose generator is monitored during the simulation to determine whether the command sequence needs to be corrected.

[0105] The actuator's actions and the speed trajectory need to be precisely coordinated in time. If their timing is misaligned, the speed regulation effect will deviate from expectations. Dynamic time warping can flexibly match asynchronous time sequences, making it suitable for timing verification under complex working conditions. The timestamps of the actuator's actions and the key time nodes of the speed trajectory are extracted and input into the dynamic time warping algorithm. The algorithm finds the best matching path by stretching or compressing one of the time axes. If the total deviation of the matching path is greater than a preset deviation threshold, it is marked as a timing mismatch, and the time period with the largest deviation is output. This accurately identifies the specific time period of timing mismatch, solves the speed regulation problem caused by timing mismatch, and provides a specific deviation time period reference for subsequent correction. If the timing mismatch is severe, the time synchronization problem should be corrected first before constraint boundary verification, otherwise the accuracy of the over-limit judgment will be affected.

[0106] Actuator actions or rotational speed trajectories may exceed elastic constraint boundaries. The causes of these exceedances are varied and require targeted analysis for effective correction. Simply determining whether an exceedance has occurred without tracing the underlying cause leads to blind corrective measures. By referencing the elastic constraint boundaries and comparing the actuator actions and rotational speed trajectories with their corresponding boundaries, if the actuator actions exceed the limits but the rotational speed trajectories are normal, it's possible that the actuator parameters are incorrect. If both exceed the limits simultaneously and the timing mismatch is low, it's determined to be due to timing misalignment. If the exceedances persist even with timing matching, it's attributed to a problem with the parameter design of the initial command sequence. This precise identification of the cause of the exceedances avoids blind corrections and ensures that subsequent adjustments directly address the root cause of the problem. Clearly defined causes guide the key monitoring directions of the simulation, making the simulation more targeted.

[0107] Even if the timing is matched and the limits are not exceeded, the instruction sequence may still cause potential damage to mechanical components. Simulation execution can detect such hidden problems in advance and avoid equipment wear and tear during actual operation. The instruction sequence is simulated and executed virtually by combining the dynamic characteristic parameters of the target prime mover. During the simulation, the mechanical state is monitored. If the bearing temperature shows an abnormal upward trend or the characteristic frequency of wear appears in the vibration spectrum, the instruction sequence is determined to need to be corrected. This allows for the early identification of potential hazards of the instruction sequence to mechanical components, making up for the limitations of relying solely on constraint boundary verification and ensuring the rationality of the instruction sequence in terms of safety and equipment protection. The simulation results determine whether to return to the target optimization module to correct the instruction sequence. If there is a risk to the mechanical state, the instruction sequence regeneration process is triggered. If there are no errors, the execution gain adjustment stage is entered.

[0108] Furthermore, the gain adjustment sub-logic includes:

[0109] The load change rate of the general generator is acquired in real time, and the correlation analysis between the load change rate and the speed fluctuation of the target prime mover is performed to generate a load-speed correlation map.

[0110] Based on the load-speed correlation graph, the execution gain of the command sequence is determined by combining the torque response delay of the target prime mover, and the command sequence is adjusted based on the execution gain;

[0111] The execution status of the instruction sequence is continuously monitored, and the difference between the actual execution effect and the expected execution effect is calculated to determine whether the execution gain of the instruction sequence should be adjusted again in combination with the vibration frequency.

[0112] Load change rate is a key factor affecting speed stability. The speed fluctuation of the target prime mover determines its ability to cope with load changes. The correlation between the two is the basis for determining the execution gain. Ignoring this correlation will make the gain adjustment lack specificity. The load change rate and the corresponding speed fluctuation are acquired in real time, and a correlation graph is drawn between the two using correlation analysis tools to obtain the load-speed correlation graph. In the load-speed correlation graph, nodes represent different load change rates and speed fluctuations, and the thickness of the edges represents the correlation strength between the two. This intuitively presents the intrinsic relationship between load change rate and speed fluctuation, providing a data-driven correlation basis for execution gain adjustment, so that the gain setting can match the specific load-speed interaction mode. The load-speed correlation graph provides a scenario classification basis for determining the execution gain in combination with torque response delay. Different correlation modes correspond to different gain adjustment strategies.

[0113] The execution gain determines the intensity of the command sequence execution. Prime movers with long torque response delays require higher gains to accelerate the response speed, but excessive gain can lead to overshoot. Dynamic balancing based on the load-speed correlation graph is necessary. Based on the load-speed correlation graph, for modes with high correlation intensity, the execution gain is adjusted in conjunction with the torque response delay. For long torque response delays, the execution gain is increased to compensate for lag, while for short torque response delays, the gain is decreased to avoid overshoot. The determined execution gain is applied to the command sequence to adjust the amplitude and rate of the actuator's action. This allows the execution gain to simultaneously adapt to the load change requirements and the dynamic response capability of the prime mover, ensuring timely speed regulation while avoiding system oscillations caused by improper gain. The initially adjusted command sequence enters the actual execution stage, providing basic execution effect data for subsequent secondary adjustments based on vibration frequency.

[0114] In actual execution, even if the gain is adapted to the load and torque characteristics, abnormal vibration may still occur due to mechanical resonance. Vibration frequency is a key indicator for monitoring this hidden problem, and the gain needs to be optimized again based on this. The vibration frequency during the execution of the instruction sequence is continuously monitored. If a vibration component close to the inherent frequency of the equipment appears, and the deviation between the actual execution effect and the expected execution effect exceeds the acceptable range, the execution gain of the corresponding frequency range is reduced. If the vibration is normal but the execution effect is deviated, the amplitude of the execution gain is finely adjusted rather than the frequency. In this way, the risk of mechanical resonance is eliminated through feedback correction of vibration frequency, and the stability of the execution effect is further improved. The gain adjustment takes into account both dynamic response and mechanical safety. Finally, the adjusted instruction sequence enters the stable execution stage, and its execution state will be used for feedback update of the perception state diagram to form a closed-loop optimization.

[0115] Specifically, such as Figure 4 As shown, the feedback update logic of the perception state diagram includes:

[0116] The parameter data of the general generator during the execution of the instruction sequence is obtained, and the parameter data is compared with the local state features in the perception state diagram to calculate the feature deviation.

[0117] Configure a deviation threshold. When the feature deviation is less than the deviation threshold, replace the node with the local state feature in the perception state graph, and update the connection weight between the nodes according to the execution state of the instruction sequence.

[0118] When the characteristic deviation is greater than or equal to the deviation threshold, the cause of the deviation is analyzed by combining the actuator action and the load change rate of the general generator, so as to determine whether to add a new directional edge in the perception state diagram.

[0119] The updated node and edge connection weights are time-series aligned with historical power generation data to verify the rationality of the perception state graph update.

[0120] After the command sequence is executed, the actual parameter data of the general-purpose generator will deviate from the local state features in the perception state diagram. This deviation reflects the accuracy of the perception state diagram in describing the current operating condition and is a prerequisite for updating the perception state diagram. The parameter data during the execution process is acquired and compared item by item with the local state features of the corresponding time window in the perception state diagram. The deviation of each parameter data from the local state features is determined, including numerical deviation and trend of change, and a feature deviation is formed. This quantifies the difference between the perception state diagram and the actual operating condition, provides a clear trigger signal for updating the perception state diagram, avoids blind updates or omission of necessary adjustments, and the feature deviation result determines the update method, providing a data basis for subsequent deviation threshold judgment.

[0121] When the feature deviation is small, it indicates that the overall perception state diagram is accurate, and only the node features need to be updated. When the feature deviation is large, there may be new relationships that have not been captured, and new directional edges need to be added to improve the structure of the perception state diagram. If the deviation is not differentiated and updated uniformly, it will lead to low efficiency or structural redundancy. When the feature deviation is less than the deviation threshold, the local state features of the corresponding node are replaced with actual parameter data, and the connection weights of the edges between nodes are updated according to the execution status. When the feature deviation is greater than or equal to the deviation threshold, the cause of the deviation is analyzed by combining the actuator action and the load change rate of the general generator. If a new causal relationship is found, a new directional edge is added between the corresponding nodes. This ensures the timeliness of the perception state diagram and expands the ability to capture new relationships, enabling the perception state diagram to dynamically adapt to changes in operating conditions. The updated nodes and edges provide new execution status and correlation information for the time-series alignment verification of historical power generation data, ensuring the accuracy of the verification basis.

[0122] The updated perception state graph needs to be compared with historical power generation data to verify whether it conforms to the long-term operating rules of the system and avoid erroneous updates caused by short-term abnormal data. The connection weights of the updated nodes and edges are time-series aligned with historical power generation data. By comparing the historical perception state graphs under similar operating conditions, it is determined whether the newly added directional edges have similar correlations in the historical power generation data, and whether the local state characteristics of the updated nodes are consistent with historical trends. If they conform to historical patterns, the update is confirmed to be effective; otherwise, it is marked as pending observation and not included in subsequent decisions. Through verification with historical power generation data, invalid updates caused by short-term anomalies are filtered out to ensure that the update of the perception state graph conforms to the inherent operating rules of the system and improves its reliability as input for subsequent modules. The verified perception state graph is fed back to the global perception module as the basis for the construction of the next round of perception state graphs, forming a closed loop of perception, optimization, execution, and feedback.

Claims

1. An intelligent speed-regulating universal generator system, characterized in that, include: The module consists of a global perception module, a working condition prediction module, a target optimization module, and an execution scheduling module. The global perception module is used to receive the trigger signal of the target prime mover connected to the general generator, acquire and fuse the static and dynamic characteristic parameters of the target prime mover and the parameter data of the general generator in real time to form fused power generation data, and align the fused power generation data with historical power generation data in time sequence to construct a perception state diagram. The operating condition prediction module is used to process the changing trends of fused power generation data through a long short-term memory network to deduce the evolution trend of operating conditions. The target optimization module is used to receive the static characteristic parameters of the target prime mover to determine the operating constraints of the target prime mover, analyze the user-preset optimization target, and generate an instruction sequence including speed trajectory and actuator action based on the perception state diagram and the working condition evolution trend through reinforcement learning. The execution scheduling module is used to receive and verify the instruction sequence, and dynamically adjust the execution gain of the instruction sequence according to the load change rate of the general generator and the dynamic characteristic parameters of the target prime mover, and monitor the execution status of the instruction sequence to update the perception status map.

2. The intelligent speed-regulating universal generator system as described in claim 1, characterized in that, The construction logic of the perception state diagram includes: The merged power generation data is time-series aligned with historical power generation data, and non-uniform time series interpolation is performed on the merged power generation data through dynamic time warping. The fused power generation data is divided by adaptive time windows to extract local state features. The local state features within each time window constitute the nodes of the perception state map. The cosine similarity and dynamic time bending distance of adjacent nodes are calculated to determine the edges and connection weights of adjacent nodes. Oriented edges of non-adjacent nodes are added through a causal inference algorithm to construct a perceptual state graph.

3. The intelligent speed-regulating universal generator system as described in claim 2, characterized in that, The sub-logic for forming the fused power generation data includes: The static characteristic parameters of the target prime mover are characterized and anchored, and the dynamic characteristic parameters of the target prime mover and the parameter data of the general generator are subjected to multi-mode filtering. The data type is determined and the correlation of data of the same type is aggregated. At the same time, the data of different types are fused across domains through a heterogeneous graph neural network. The static characteristic parameters of the target prime mover are used as the initial node features of the heterogeneous graph neural network. The dynamic characteristic parameters of the target prime mover and the parameter data of the general generator are used as edge features, and the interaction weights of different types of data are learned through an attention mechanism to form fused power generation data.

4. The intelligent speed-regulating universal generator system as described in claim 3, characterized in that, The deductive logic for the evolution trend of the operating conditions includes: The integrated power generation data is divided into high-frequency and low-frequency components according to the time dimension, and the correlation between the integrated power generation data is sorted out according to the spatial dimension. The integrated power generation data is then calculated by mutual information entropy to filter out strongly correlated data pairs. Long Short-Term Memory (LSTM) networks consist of short-term, medium-term, and long-term branches. High-frequency components are input to the short-term branch to output short-term prediction sequences, low-frequency components and short-term prediction sequences are input to the medium-term branch to output phase prediction sequences, and strongly correlated data pairs and phase prediction sequences are input to the long-term branch to output trend prediction sequences. The branch weights are dynamically adjusted according to the complexity of the working conditions. After fusion, a preliminary trend sequence is formed. The perception state map is received and the correlation strength between nodes is calculated through the graph attention network to generate a topology vector. The preliminary trend sequence and the topology vector are then fused to generate a trend matrix. By processing the trend matrix and historical prediction errors through Bayesian inference, the confidence interval of each prediction result is calculated. At the same time, the confidence interval of the prediction result is corrected according to the degree centrality of the nodes in the perception state graph to infer the evolution trend of the working condition. The difference between the evolution trend of the working condition and the actual data is used to determine whether to adjust the branch weights of the long short-term memory network.

5. The intelligent speed-regulating universal generator system as described in claim 4, characterized in that, The logic for generating the instruction sequence includes: The local state features and edge connection weights in the perception state graph are used to generate an initial state vector, and then the evolution trend weights of the working conditions are assigned through an attention mechanism and fused into a state vector. The algorithm analyzes user-defined optimization goals and generates basic rewards, and applies negative rewards to actions that violate operational constraints in order to determine the reward function for reinforcement learning. Based on the state vector and reward function, candidate speed trajectories are generated through Monte Carlo tree search. The candidate speed trajectories are verified by running constraints to determine the speed trajectory. Based on the speed trajectory, the executor actions are generated through the near-end policy algorithm to generate the instruction sequence.

6. The intelligent speed-regulating universal generator system as described in claim 5, characterized in that, The sub-logic for determining the operational constraints includes: Receive the static characteristic parameters of the target prime mover, set a multi-dimensional constraint space based on the rated power, efficiency curve and structural parameters, and generate initial constraint conditions in combination with the type of the target prime mover; Based on historical operating data, elastic constraint boundaries are set for each initial constraint condition, and the confidence level of the elastic constraint boundaries is calculated through Bayesian inference. The elastic constraint boundary is dynamically adjusted according to the evolution trend of the operating conditions in order to determine the operating constraints of the target prime mover.

7. The intelligent speed-regulating universal generator system as described in claim 6, characterized in that, The feedback update logic of the perception state diagram includes: The parameter data of the general generator during the execution of the instruction sequence is obtained, and the parameter data is compared with the local state features in the perception state diagram to calculate the feature deviation. Configure a deviation threshold. When the feature deviation is less than the deviation threshold, replace the node with the local state feature in the perception state graph, and update the connection weight between the nodes according to the execution state of the instruction sequence. When the characteristic deviation is greater than or equal to the deviation threshold, the cause of the deviation is analyzed by combining the actuator action and the load change rate of the general generator, so as to determine whether to add a new directional edge in the perception state diagram. The updated node and edge connection weights are time-series aligned with historical power generation data to verify the rationality of the perception state graph update.

8. The intelligent speed-regulating universal generator system as described in claim 7, characterized in that, The verification logic of the instruction sequence includes: Receive instruction sequences, extract timestamps of actuator actions and time nodes of rotational speed trajectories, and verify the timing matching degree between actuator actions and rotational speed trajectories through dynamic time warping; Call the elastic constraint boundary in the running constraints, and compare the actuator action and speed trajectory with the elastic constraint boundary to determine the cause of the over-limit; By combining the dynamic characteristic parameters of the target prime mover, the execution of the command sequence is simulated, and the mechanical state of the general-purpose generator is monitored during the simulation to determine whether the command sequence needs to be corrected.

9. The intelligent speed-regulating universal generator system as described in claim 8, characterized in that, The adjustment sub-logic for the execution gain includes: The load change rate of the general generator is acquired in real time, and the correlation analysis between the load change rate and the speed fluctuation of the target prime mover is performed to generate a load-speed correlation map. Based on the load-speed correlation graph, the execution gain of the command sequence is determined by combining the torque response delay of the target prime mover, and the command sequence is adjusted based on the execution gain; The execution status of the instruction sequence is continuously monitored, and the difference between the actual execution effect and the expected execution effect is calculated to determine whether the execution gain of the instruction sequence should be adjusted again in combination with the vibration frequency.

10. The intelligent speed-regulating universal generator system as described in claim 9, characterized in that, The static characteristic parameters of the target prime mover include rated power, efficiency curve, and structural parameters. The dynamic characteristic parameters of the target prime mover include speed fluctuation, torque response delay, and vibration frequency. The parameter data of the general-purpose generator include electrical parameters, mechanical status, environmental information, and load characteristics. The electrical parameters of the general-purpose generator include generator speed, voltage harmonic distortion rate, and power factor. The mechanical status of the general-purpose generator includes bearing temperature and vibration spectrum. The environmental information of the general-purpose generator includes ambient temperature and ambient humidity. The load characteristics of the general-purpose generator include load change rate and power distribution timing.

Images