A load prediction-based dynamic regulation method for a turbine power generation propulsion system

Abstract

The application discloses a kind of based on load prediction turbine power generation propulsion system dynamic regulation method, comprising: the operating state information in flight task is collected and power consumption record;Information is input to trained data processing unit, and the expected power demand of each power consumption node is generated;According to the matching relationship of expected demand and generator current output, it is judged whether there is high load fluctuation sign;If there is, then with expected demand as constraint, the power transmission path is compared by optimization algorithm to multiple schemes iteration, and the optimal power allocation strategy is determined;According to optimal strategy adjustment generator set output parameter and distribution passage state, and continuously collect feedback data;When feedback data shows that actual state deviates from expected target, according to the deviation degree, expected demand is regenerated and returned to strategy determination step of execution.The application realizes prospective regulation by load prediction, identifies fluctuation risk and dynamically optimizes power allocation, and combines feedback iteration to correct deviation.

Description

Technical Field

[0001] This invention relates to the field of control technology for turbine-powered propulsion systems, and in particular to a dynamic control method for turbine-powered propulsion systems based on load prediction. Background Technology

[0002] In the field of aerospace technology, the development of turbine-generated propulsion systems has always held a core strategic position, as their performance directly determines the power output quality and mission reliability of aircraft. With the deepening development of the more-electric aircraft concept, modern aircraft are increasingly reliant on electrical systems. From the main propulsion motor to avionics, from environmental control systems to actuators, the normal operation of almost every critical functional component relies on a stable and sufficient power supply. However, flight missions are inherently dynamic processes. The rapid power surge during climb, the relative stability of cruise, and the sudden loads during maneuvers all place stringent demands on the responsiveness and power distribution logic of the electrical system.

[0003] Most power management solutions currently used in the industry are based on logic control using preset rules. This means that the generator's output baseline value is pre-set according to the flight phase, and feedback regulation is performed by collecting real-time voltage and current signals. While this control architecture can meet basic power supply needs under steady-state conditions, it is inherently a delayed response – the control mechanism only intervenes after power fluctuations have occurred and system parameters have deviated. This reactive characteristic often proves inadequate when facing drastically changing power demands during flight. For example, when an aircraft performs an emergency climb or evasive maneuver, the power demand of the propulsion motor may increase exponentially within seconds. Traditional feedback control systems, unable to anticipate this change, can only react passively, easily leading to voltage drops, frequency fluctuations, or even protective shutdowns of equipment, directly impacting flight safety.

[0004] To address the aforementioned issues, some cutting-edge research has attempted to introduce a segmented pre-set method based on flight plans. This involves switching different power generation control parameters at different stages according to a pre-set flight profile. While this approach improves the system's adaptability to some extent, it remains essentially an open-loop feedforward control system, unable to handle unexpected situations or mission changes during actual flight. Once the aircraft deviates from the pre-set route or encounters unexpected weather conditions, the pre-set parameters become meaningless, and the system still falls into passive, lagging adjustment. More importantly, this type of approach lacks the ability to dynamically balance the priority relationships between different power-consuming devices, making it difficult to achieve optimal allocation of scarce power resources when total power generation is limited. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the existing power management system, which adopts ex-post feedback regulation or open-loop preset method, making it difficult to cope with the drastic fluctuations in power demand during flight, resulting in system response lag, poor power supply stability and insufficient power supply guarantee for key equipment. The present invention provides a dynamic control method for turbine generator propulsion system based on load prediction. It achieves forward-looking regulation through load prediction, identifies fluctuation risks based on supply and demand matching relationship, and dynamically optimizes power allocation through optimization algorithm when risks occur. It iteratively corrects prediction deviations by combining feedback data, thereby effectively improving the stability of the power supply system, the power supply guarantee rate of key equipment, and the system's adaptability to changes in flight conditions.

[0006] To address the aforementioned technical problems, this invention provides a dynamic control method for a turbine-powered propulsion system based on load prediction, comprising the following steps:

[0007] Collect operational status information and corresponding power consumption records during flight missions;

[0008] The operating status information and power consumption records are input into the trained data processing unit to generate the expected power demand of each power consumption node in the future period;

[0009] Based on the matching relationship between expected electricity demand and the current output capacity of generators, determine whether there are signs of high load fluctuations;

[0010] If there are signs of high load fluctuations, the optimal power dispatching strategy is determined by iteratively comparing multiple power transmission paths using an optimization algorithm, with the expected power demand as a constraint.

[0011] Based on the optimal power dispatch strategy, adjust the output parameters of the generator sets and the on / off status of the power distribution channels, and continuously collect system operation feedback data;

[0012] When feedback data shows that the actual operating status of the system deviates from the expected allocation target, the expected power demand is regenerated based on the degree of deviation, and the process returns to the step of determining the optimal power allocation strategy.

[0013] In one embodiment of the present invention, the operational status information includes at least one of flight altitude, flight speed, rate of climb, pitch angle, and ambient temperature.

[0014] In one embodiment of the present invention, the trained data processing unit is constructed in the following manner:

[0015] Extract no less than 1,000 sets of historical flight mission data from the flight data recorder. Each set of data contains the operational status information at the same moment and the corresponding power consumption records of each power consumption node.

[0016] Using operational status information as input samples and power consumption records as output labels, the multilayer feedforward neural network is iteratively trained using the backpropagation algorithm.

[0017] During training, the input samples are propagated forward layer by layer to calculate the network output. The mean square error between the network output and the output label is used as the loss function. The partial derivative of the loss function with respect to the weights of each layer is used as the error signal and propagated backward to update the network weights layer by layer.

[0018] Repeat the iteration until the loss function value converges to a preset range, and obtain a data processing unit that can output the corresponding power consumption based on the input operating status information.

[0019] In one embodiment of the present invention, the expected power demand of each power-consuming node in the future time period is represented in the form of power-time series. Specifically, the generation method is as follows: the operating status information of the current moment and the previous 10 consecutive sampling points are combined with the power consumption record to form an input vector, which is input to the trained data processing unit, and the power demand value of each power-consuming node is output every 0.5 seconds in the next 5 to 30 seconds, forming a power demand matrix of 60 time points in the future time period, including four power-consuming nodes: propulsion motor, environmental control system, avionics equipment, and actuation mechanism.

[0020] In one embodiment of the present invention, determining whether there are signs of high load fluctuations based on the matching relationship between expected power demand and the current output capacity of the generator specifically includes:

[0021] The expected power demand of each power consumption node in the future period is summed at the same time to obtain the total expected power demand sequence at each time point in the future period.

[0022] Collect the current maximum output power of the generator and compare it with the maximum value in the total expected power demand sequence; calculate the difference between the maximum value and the current maximum output power of the generator. If the difference is less than 15% of the generator's rated power, it is determined that there is a sign of high load fluctuation. At the same time, calculate the power change rate at adjacent time points in the total expected power demand sequence. If there are three or more consecutive time points where the power change rate exceeds 20% per second, it is also determined that there is a sign of high load fluctuation.

[0023] In one embodiment of the present invention, the optimization algorithm is a particle swarm optimization algorithm, which performs multiple schemes for power transmission path iterative comparison through the following steps:

[0024] Initialize a population of 50 particles, each with an eight-dimensional position vector, corresponding to the power allocation ratio of the four main power-consuming nodes and the switching state of the four key power distribution paths.

[0025] Each particle is randomly initialized with a position and velocity. The current position of each particle is taken as its individual historical best position, and the position with the best fitness value among all particles is taken as the group's historical best position.

[0026] The fitness function is defined as the sum of the squares of the deviations between the actual power allocated to each power consumption node and the expected power demand, plus a penalty term for the number of switching times of the distribution path.

[0027] Each particle iterates according to its own historical best position and the group's historical best position, following the velocity update formula and the position update formula. The velocity update formula includes inertia weight, cognitive learning factor, and social learning factor.

[0028] After each iteration, the fitness value of each particle is recalculated, and the individual historical best position and the group historical best position are updated.

[0029] Repeat the iteration until the preset number of iterations of 200 is reached or the group's historical best position remains unchanged for 30 consecutive iterations. Then, decode the eight-dimensional vector corresponding to the final group's historical best position into the power allocation ratio of each power consumption node and the switching state of the power distribution path, which will serve as the optimal power dispatching strategy.

[0030] In one embodiment of the present invention, the position vector encoding method for each particle in the particle swarm optimization algorithm is as follows:

[0031] The first four dimensions of the position vector correspond to the power distribution ratio of the propulsion motor, environmental control system, avionics equipment, and actuator, respectively. The values ​​of the first four dimensions range from 0 to 1, and the sum of the four is 1. This is achieved by performing softmax normalization on the original position values.

[0032] The last four dimensions of the position vector correspond to the switching states of the first power distribution path connecting the generator and the propulsion motor, the second power distribution path connecting the generator and the environmental control system, the third power distribution path connecting the generator and the avionics equipment, and the fourth power distribution path connecting the generator and the actuation mechanism. The values ​​of the last four dimensions are mapped to the interval between 0 and 1 by the sigmoid function, and then binarized to 0 or 1 with a threshold of 0.5, where 0 represents disconnection and 1 represents connection.

[0033] In one embodiment of the present invention, adjusting the generator set output parameters and the on / off state of the power distribution path specifically includes:

[0034] The generator electronic control unit sends a target output power command, which includes the power ramp-up rate limit and steady-state output power value for the next 5 seconds. The generator electronic control unit adjusts the excitation current and fuel flow according to the command, so that the generator output power smoothly transitions to the target value at the specified ramp-up rate.

[0035] Simultaneously, control commands are sent to the solid-state power controller array in the power distribution network. Each solid-state power controller corresponds to a power consumption node or power distribution path. The control commands include on / off status and target current limit value. After receiving the on command, the solid-state power controller closes its internal power switch under the condition of not exceeding the target current limit value, thus completing the reconfiguration of the power transmission path.

[0036] In one embodiment of the present invention, the system operation feedback data includes at least one of the following: real-time voltage, real-time current, real-time power, and generator output frequency of each power consumption node.

[0037] In one embodiment of the present invention, when feedback data shows a deviation between the actual operating status of the system and the expected dispatch target, the expected power demand is regenerated based on the degree of deviation, specifically including:

[0038] The real-time power values ​​of each power consumption node are collected and subtracted point by point from the expected power allocation value in the optimal power dispatch strategy to obtain the real-time power deviation sequence of each power consumption node.

[0039] The mean of the real-time power deviation sequence of each power consumption node within the next second is calculated as the steady-state deviation, and the variance of the sequence is calculated as the fluctuation deviation.

[0040] The steady-state deviation and the fluctuation deviation are multiplied by the first correction coefficient and the second correction coefficient, respectively, and then superimposed on the input vector of the data processing unit in the next cycle. The first correction coefficient is used to compensate for systematic prediction deviation, and its value is positively correlated with the degree of influence of steady-state deviation on prediction accuracy. The second correction coefficient is used to suppress prediction fluctuation error, and its value is positively correlated with the degree of influence of fluctuation deviation on system stability. The corrected input vector is input to the trained data processing unit to regenerate the expected power demand of each power consumption node in the future period.

[0041] The technical solution of the present invention has the following advantages compared with the prior art:

[0042] The dynamic control method for turbine-powered propulsion systems based on load forecasting described in this invention effectively overcomes the limitations of traditional delayed responses by proactively predicting future power demand. The method identifies signs of high load fluctuations before they occur and dynamically optimizes power allocation strategies accordingly, ensuring a stable and reliable power supply during flight missions, even in the face of drastically changing power demands. Furthermore, the feedback correction mechanism introduced in the method further enhances the system's adaptability to deviations from actual operation, helping to maintain stable power system operation, avoid voltage drops and frequency fluctuations, and thus improve the reliability of the aircraft's mission execution. Attached Figure Description

[0043] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0044] Figure 1 This is a flowchart of the steps of the dynamic control method for a turbine-powered propulsion system based on load prediction according to the present invention;

[0045] Figure 2 This is a flowchart of the steps for constructing a trained data processing unit according to the present invention;

[0046] Figure 3 This is a flowchart of the particle swarm optimization algorithm of the present invention;

[0047] Figure 4 This is a flowchart of the steps in this invention to adjust the output parameters of the generator set and the on / off state of the power distribution path;

[0048] Figure 5 This is a flowchart illustrating the steps of regenerating the expected electricity demand based on the degree of deviation in this invention. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0050] Reference Figure 1 As shown, this invention proposes a dynamic control method for a turbine-powered propulsion system based on load prediction. The method first collects operational status information and corresponding power consumption records during flight missions, inputting this real-time data into a trained data processing unit to generate the expected power demand for each power consumption node in the future. This step shifts the control perspective from reactive to proactive—the system no longer waits for power fluctuations to occur before taking remedial action, but rather anticipates upcoming load changes based on the current flight situation. The trained data processing unit can learn the inherent correlation between different flight states and power demand from historical operational data, thereby achieving accurate prediction of future demand and gaining a valuable time window for subsequent control.

[0051] Based on the anticipated electricity demand, the system further determines whether there are signs of high load fluctuations by analyzing the matching relationship between the anticipated electricity demand and the current output capacity of the generators. This step essentially involves assessing the power supply and demand situation and identifying potential risks of supply-demand imbalance. When signs of high load fluctuations are detected, it indicates that the current generation capacity will soon be unable to meet the upcoming load demand, and the system immediately enters active optimization mode.

[0052] In the active optimization mode, with expected power demand as a constraint, an optimization algorithm iteratively compares multiple power transmission path options to ultimately determine the optimal power allocation strategy. This optimization algorithm can search for the most reasonable power allocation scheme in complex distribution networks. Its decision-making process comprehensively considers multiple factors such as the priority of each power-consuming node, line transmission capacity, and switch status limitations, ensuring that critical equipment receives priority power supply when total power generation is limited, while non-critical equipment is appropriately deloaded or has its power supply delayed. Compared to traditional fixed rules or simple priority-based power cutoffs, this optimization-based allocation method maximizes the utility of power resources.

[0053] After determining the optimal strategy, the generator output parameters and the on / off status of the power distribution path are adjusted accordingly to implement the optimized plan, while continuously collecting system operation feedback data. This step transforms the optimization results from the decision-making layer into actual control actions at the physical layer, and the continuous collection of feedback data provides an information foundation for closed-loop control. When feedback data shows a deviation between the actual system operating status and the expected dispatch target, the method can regenerate the expected power demand based on the degree of deviation and return to the step of determining the optimal power dispatch strategy, forming a dynamic iterative control closed loop.

[0054] From a fundamental perspective, the reason why the technical solution adopted in this invention can solve the problems existing in the background technology lies in its construction of a complete control loop of "prediction-evaluation-optimization-execution-feedback". Compared with traditional solutions, its essential difference lies in the introduction of a load prediction mechanism, which enables the system to have a forward-looking adjustment capability, thereby transforming the response mode from post-event correction to pre-event prevention. On this basis, dynamic resource allocation is achieved through optimization algorithms, overcoming the limitation that fixed rules cannot adapt to changing operating conditions. The introduction of a feedback iteration mechanism ensures that the system can continuously correct deviations during continuous operation and adapt to the dynamic changes of flight missions.

[0055] Based on the above technical principles, the beneficial effects that this invention can achieve include at least the following:

[0056] By introducing a load forecasting mechanism, the system can anticipate changes in power demand and adjust power generation output and distribution configuration before load surges occur, thereby effectively avoiding power quality problems such as voltage drops and frequency fluctuations and significantly improving the stability of the power supply system.

[0057] By using an optimization algorithm to iteratively compare multiple power transmission paths, the system can dynamically allocate power based on the real-time importance of each electrical device under conditions of limited power resources. This ensures that critical flight safety equipment receives priority support under any operating condition, thereby improving the reliability of mission execution.

[0058] The continuous collection of feedback data and iterative correction of deviation enable the system to have adaptive learning capabilities. As operational experience accumulates, the system can continuously optimize prediction accuracy and the applicability of allocation strategies, maintaining excellent control performance under different flight missions and environmental conditions.

[0059] In summary, this invention provides a dynamic control solution for turbine power generation propulsion systems that is both forward-looking and adaptable through the deep integration of prediction and optimization, effectively improving the stability, reliability and intelligence level of aviation power systems.

[0060] In this embodiment, the operational status information includes at least one of flight altitude, flight speed, rate of climb, pitch angle, and ambient temperature. Specifically, the operational status information is the basic input data for the system to predict power demand. These parameters directly reflect the current motion state of the aircraft and its environment, and are closely related to the load demand of the turbine-generator propulsion system. For example, flight altitude and flight speed directly affect the air resistance experienced by the aircraft, thus affecting the power demand of the propulsion motor; rate of climb and pitch angle reflect the attitude changes and vertical motion trends of the aircraft, which may lead to changes in the load of the actuators and environmental control system; ambient temperature directly affects the power consumption of the environmental control system (such as air conditioning and cooling systems). Therefore, selecting these key parameters as operational status information ensures a strong physical correlation between the input data and power demand.

[0061] In practice, flight altitude can be obtained through an onboard atmospheric data computer (ADC) or Global Positioning System (GPS), typically measured in meters or feet. Flight speed can be calculated using an airspeed tube, pitot tube, and ADC, or obtained from GPS ground speed data, typically measured in meters per second or knots. Rate of climb can be calculated using a vertical velocity sensor or ADC, representing the change in the aircraft's vertical velocity, typically measured in meters per second or feet per minute. Pitch angle can be obtained through an inertial measurement unit (IMU) or Attitude and Heading Reference System (AHRS), representing the angle of the aircraft's nose relative to the horizontal plane. Ambient temperature can be obtained through temperature sensors installed externally to the aircraft, typically measured in degrees Celsius or Kelvin. The acquisition of these data should be highly accurate and real-time to ensure that the input data accurately reflects the instantaneous state of the aircraft. Sensors should be rigorously calibrated and possess appropriate anti-interference capabilities to guarantee data quality. The data acquisition frequency should be matched with the sampling frequency of the data processing unit to avoid information lag or loss, thus providing a reliable basis for accurate predictions.

[0062] The above method requires a trained data processing unit to accurately generate the expected power demand of each power consumption node in the future. However, how to effectively construct and train this data processing unit to accurately capture the nonlinear relationship between the complex operating states and power consumption during flight missions is the key to ensuring prediction accuracy and system control effectiveness.

[0063] In this regard, refer to Figure 2 As shown, this application further proposes a method for constructing the trained data processing unit, specifically including:

[0064] To construct a data processing unit capable of accurately predicting power demand, sufficient and representative historical data is first required. This application achieves this by extracting no fewer than 1000 sets of historical flight mission data from flight data recorders. Each set of data includes operational status information at the same moment and corresponding power consumption records for each power-consuming node. This historical data is the cornerstone of training the prediction model, and its quantity and quality directly affect the model's generalization ability and prediction accuracy. For example, operational status information may include flight altitude, flight speed, rate of climb, pitch angle, ambient temperature, etc., while power consumption records detail the real-time power consumption of key power-consuming nodes such as propulsion motors, environmental control systems, avionics equipment, and actuators. By collecting sufficient historical data, various operating conditions that may occur during flight missions can be covered, thereby enabling the trained model to have stronger adaptability.

[0065] After acquiring historical data, this application uses operational status information as input samples and power consumption records as output labels, employing an error backpropagation algorithm to iteratively train a multilayer feedforward neural network. A multilayer feedforward neural network is a machine learning model with powerful nonlinear mapping capabilities, capable of learning the complex relationship between input (operational status information) and output (power consumption records). The error backpropagation algorithm is a common and efficient method for training such neural networks. It calculates the output error and propagates it back through each layer of the network, thereby adjusting the weights and biases within the network, gradually bringing the network's output closer to the true label.

[0066] During training, the input samples are first propagated forward layer by layer, and the network output is calculated by the neurons in each layer. Then, this network output is compared with the corresponding output label (i.e., the actual electricity consumption record), and the mean squared error (MSE) between the two is calculated as the loss function. The MSE quantifies the difference between the network's prediction and the actual value. Next, using the chain rule, the partial derivatives of the loss function with respect to the weights and biases of each layer in the network are calculated; these partial derivatives are the error signals. These error signals propagate from the output layer to the input layer, guiding the update of the network weights in each layer. In this way, the network can gradually adjust its internal parameters according to the magnitude and direction of the error to reduce the prediction error.

[0067] The above training process is repeated iteratively until the loss function value converges to a preset range. This preset range is manually set to indicate that the network's prediction error is sufficiently small, and the model has learned a stable mapping relationship between input and output. Once the convergence condition is met, a data processing unit is obtained that can accurately output the corresponding power consumption based on the input operating status information. This data processing unit has the ability to predict future power demand, providing a reliable basis for subsequent dynamic regulation.

[0068] In actual dynamic control processes, if the expected power demand is expressed only as a single value or a rough point in time, it may not be able to accurately capture the complex load fluctuations that may occur in the system in a short period of time, thereby affecting the accuracy and timeliness of subsequent dispatch strategies and making it difficult to effectively cope with the rapidly changing power demand of the turbine power generation propulsion system.

[0069] To this end, this application further proposes to characterize the expected electricity demand of each electricity consumption node in the future period in the form of power-time series. The specific generation method is as follows:

[0070] The anticipated electricity demand is represented in the form of a power-time series, meaning that the forecast result is no longer a single instantaneous power value, but includes a series of power predictions at different future points in time. This series format can more comprehensively reflect the load change trends and dynamic characteristics of each electricity consumption node over a period of time, providing richer and more forward-looking information for subsequent dynamic regulation. For example, it can be predicted that a certain electricity consumption node will experience a power peak at a certain time in the future, or that power will continue to rise or fall over a certain period of time.

[0071] To generate this refined power-time series, this application combines the operational status information of the current moment and the previous 10 consecutive sampling points with power consumption records into an input vector. The operational status information may include at least one of flight altitude, flight speed, rate of climb, pitch angle, and ambient temperature, while the power consumption records contain real-time power and current data for each power consumption node. Using this historical data as input enables the trained data processing unit to learn the load variation patterns and dynamic response characteristics of the system under different operational states, thereby improving the accuracy and robustness of predictions. For example, the data processing unit can be a multi-layer feedforward neural network, trained using an error backpropagation algorithm, enabling it to extract features from historical data and make effective predictions.

[0072] The input vector is then fed into a trained data processing unit. After sufficient training, this unit can output power demand values ​​for each electricity consumption node every 0.5 seconds over the next 5 to 30 seconds, based on the input historical operating status and power consumption data. This prediction time range (5 to 30 seconds) provides the system with sufficient lead time for decision-making and response, while the 0.5-second time granularity ensures the ability to capture rapid load changes and avoids control lag caused by overly coarse prediction granularity.

[0073] Ultimately, these output power demand values ​​are organized into a power demand matrix comprising four power-consuming nodes—propulsion motors, environmental control systems, avionics, and actuators—for a total of 60 time points within a future period. This matrix clearly displays the expected power demand of each key power-consuming node every 0.5 seconds over the next 30 seconds. Its structured form facilitates subsequent optimization algorithms for processing and analysis, thereby efficiently formulating the optimal power allocation strategy.

[0074] In dynamic control methods for turbine-powered propulsion systems, the system aims to achieve proactive power management by predicting the expected power demand at various power consumption nodes over future periods. However, simply obtaining the expected power demand is insufficient to accurately determine whether the system faces potential high load risks or drastic fluctuations, which may lead to unnecessary control actions or failure to respond in a timely manner when truly needed. If signs of high load fluctuations are not accurately identified, the system may be unable to effectively warn and adjust before the actual load impact arrives, thereby affecting the system's stability and reliability.

[0075] In response, this application further proposes a specific method for determining whether there are signs of high load fluctuations based on the matching relationship between expected electricity demand and the current output capacity of generators, including:

[0076] The expected power demand of each power-consuming node within a future time period is summed at the same time to obtain the total expected power demand sequence for each time point within the future time period. This step aims to integrate the scattered power demand information for each power-consuming node into a total demand at the system level. By summing the expected power demand of all power-consuming nodes (such as propulsion motors, environmental control systems, avionics equipment, actuators, etc.) at each time point within a future time period, a "total expected power demand sequence" reflecting the overall load trend of the system can be obtained. This sequence visually demonstrates the dynamic changes in the total power demand of the system over a future period, providing basic data for subsequent load assessment. In practice, a central processing unit or a dedicated digital signal processor can perform parallel or serial summation operations to aggregate the power demand data of each node from the data processing unit.

[0077] Subsequently, the current maximum output power of the generator is collected and compared with the maximum value in the total expected power demand sequence. This step is used to assess whether the generator set, under current conditions, is capable of meeting the highest possible future load demand. The current maximum output power of the generator refers to the maximum electrical power that the generator can stably output under current operating conditions (such as speed, temperature, fuel supply, etc.). This value can be obtained in real time through the generator electronic control unit, or estimated by monitoring key generator parameters through sensors and combining them with a preset model. Comparing this value with the maximum value in the total expected power demand sequence can preliminarily determine whether the system is at risk of insufficient power margin under future peak loads.

[0078] Based on this, the difference between the maximum value and the current maximum output power of the generator is calculated. If the difference is less than 15% of the generator's rated power, it is determined that there is a sign of high load fluctuation. This judgment condition aims to identify whether the power margin of the generator set is too small. When the difference between the future peak demand and the generator's maximum output capacity, i.e., the power margin, is lower than a preset percentage (e.g., 15%) of the generator's rated power, it indicates that the system's redundancy capacity of the generator set is insufficient to cope with future peak loads, and there is a high risk of overload. Even if the generator can barely meet the peak demand, an excessively small margin also means that the system's resistance to sudden disturbances is weak, which can easily lead to voltage drops or frequency fluctuations, and is therefore considered a sign of high load fluctuation. This calculation and comparison logic can be implemented by a software module in the system controller.

[0079] Simultaneously, the power change rate at adjacent time points in the total expected power demand sequence is calculated. If the power change rate exceeds 20% per second for three or more consecutive time points, a high load fluctuation is also identified. This criterion focuses on identifying drastic load change trends, rather than just peak load. The power change rate reflects the dynamic response speed of the system load. The real-time power change rate can be obtained by calculating the ratio of the power difference between adjacent time points in the total expected power demand sequence to the time interval. If the power change rate consistently exceeds a preset threshold (e.g., 20% per second) over multiple consecutive time points (e.g., three or more), it indicates that the system load is experiencing a rapid and continuous rise or fall. Such drastic dynamic changes may cause generator sets or distribution networks to have difficulty responding quickly, thus leading to system instability. This criterion complements the previous one, together constituting a comprehensive identification of high load fluctuation symptoms. The calculation and continuity judgment of this change rate can be implemented through digital filters and state machine logic.

[0080] In some of the embodiments described above in this application, an optimization algorithm is proposed to iteratively compare multiple power transmission paths to determine the optimal power allocation strategy when there are signs of high load fluctuations. However, in practical applications, the optimization of power transmission paths is a complex nonlinear multi-objective problem involving continuous variables (such as power allocation) and discrete variables (such as switching states). Traditional optimization algorithms may face problems such as slow convergence speed, easy getting trapped in local optima, or high computational resource consumption, making it difficult to quickly and effectively find the optimal solution that meets the real-time control requirements in dynamically changing flight mission environments.

[0081] In this regard, refer to Figure 4 As shown, this application further proposes that the aforementioned optimization algorithm is specifically a particle swarm optimization algorithm, which iteratively compares multiple schemes for power transmission paths through the following steps. Particle Swarm Optimization (PSO) is a swarm intelligence-based optimization algorithm that searches for the optimal solution in a multi-dimensional search space by simulating cooperation and competition among individuals in a swarm. Compared to other optimization algorithms, PSO has the advantages of simple implementation, fewer parameters, fast convergence speed, and strong global search capability, making it particularly suitable for solving complex optimization problems involving nonlinearity, multiple peak values, and high dimensionality.

[0082] In the initial stage of the particle swarm optimization algorithm, a population of 50 particles is first initialized. Each particle represents a potential solution, and its state is described by a position vector and a velocity vector. The position vector of each particle is set to an eight-dimensional vector, which encodes the power allocation ratio of the four main power-consuming nodes in the turbine-powered propulsion system, as well as the switching states of the four key power distribution paths connecting the generator to these four power-consuming nodes. This encoding method integrates continuous power allocation decisions and discrete switching state decisions into a unified search space, enabling the algorithm to optimize both key aspects simultaneously.

[0083] After population initialization, each particle is randomly assigned an initial position and initial velocity. These initial values ​​are typically generated randomly within a predefined search space to ensure the algorithm's initial exploratory nature. Subsequently, each particle's current random position is recorded as its individual historical best position, representing the best solution found by that particle so far. Simultaneously, the position of the particle with the highest fitness value among all particles is recorded as the population's historical best position, representing the best solution found by the entire population so far. These two optimal positions will guide the particle's movement direction in subsequent iterations.

[0084] To evaluate the merits of the solution represented by each particle, this application designs a fitness function. This fitness function consists of two parts: the first part is the sum of squares of the deviations between the actual power allocated to each power-consuming node and the expected power demand, aiming to minimize the gap between actual power supply and predicted demand, ensuring the accuracy of power supply; the second part is a penalty term for the number of power distribution path switching operations, aiming to minimize frequent switching of power distribution path switches while meeting power demand, as frequent switching may lead to system instability, increased equipment wear, or energy loss. Through this composite fitness function, the algorithm can achieve a balance between meeting power demand and maintaining system stability.

[0085] In each iteration, each particle iterates according to its own historical best position and the group's historical best position, following velocity and position update formulas. The velocity update formula typically comprises three main parts: an inertia term (controlled by inertia weights), which maintains the particle's current motion trend and aids in global search; a cognitive term (controlled by cognitive learning factors), which guides the particle towards its own discovered best position, reflecting individual experience; and a social term (controlled by social learning factors), which guides the particle towards the group's overall discovered best position, reflecting group collaboration. Through the balancing effect of these factors, the particle can weigh the exploration of new regions against utilizing known best regions. The position update formula simply adds the updated velocity to the current position, thus allowing the particle to move within the search space.

[0086] After each particle completes its velocity and position update, its fitness value corresponding to the new position needs to be recalculated. If a particle's fitness value at the new position is better than its current individual best position, then the particle's individual best position is updated. Similarly, if the newly updated individual best position is better than the current swarm best position, then the swarm best position is updated. This process ensures that the algorithm can continuously track and record the best solution discovered so far.

[0087] The particle swarm optimization algorithm iterates continuously until a preset termination condition is met. This application sets two termination conditions: one is reaching a preset number of 200 iterations to control the algorithm's running time; the other is when the swarm's historical best position remains unchanged for 30 consecutive iterations, indicating that the algorithm may have converged to a stable solution, and further iterations will yield little benefit. The algorithm stops iterating when either condition is met.

[0088] When the algorithm terminates, the eight-dimensional vector corresponding to the final historical best position of the population is considered the optimal solution to the current optimization problem. This eight-dimensional vector needs to be decoded to extract the specific power allocation strategy, including the power distribution ratio of the four main power-consuming nodes and the switching states of the four key power distribution paths. These decoded parameters will be directly used to guide the actual operation and control of the turbine-powered propulsion system, forming the optimal power allocation strategy.

[0089] Specifically, the first four dimensions of the position vector characterize the power allocation ratios of the four main power-consuming nodes within the system: the propulsion motor, the environmental control system, the avionics equipment, and the actuators. These ratios collectively determine the share of electrical energy that each power-consuming node can obtain under the total available power. By optimizing these ratios, the system's operating efficiency or stability can be maximized while meeting the needs of each power-consuming node. The power allocation ratio is limited to a range of 0 to 1, representing the percentage of power obtained by each power-consuming node relative to the total power. Simultaneously, the sum of these four ratios must be 1, ensuring that all available power is fully allocated without over-allocation or under-allocation, thus guaranteeing the physical rationality of the power allocation scheme. To achieve this constraint, this application employs softmax normalization. Softmax normalization is a method that converts any real-valued vector into a probability distribution vector. Here, it is used to convert the original position values ​​generated in continuous space by the particle swarm optimization algorithm into power allocation ratios that satisfy the constraint of "values ​​ranging from 0 to 1 and the sum of the four ratios being 1".

[0090] Meanwhile, the last four dimensions of the position vector are used to characterize the switching states of the four key power distribution paths connecting the generator to the four main power-consuming nodes (propulsion motor, environmental control system, avionics equipment, and actuators). These switching states determine whether power can be delivered from the generator to the corresponding power-consuming nodes. By adjusting these switching states, flexible configuration of the power transmission paths can be achieved to cope with different load demands and system operating conditions. To convert the continuous position values ​​generated by the particle swarm optimization algorithm in the continuous space into discrete switching states (0 or 1), this application first maps the original position values ​​to the range of 0 to 1 using the sigmoid function, giving them a probabilistic meaning. The sigmoid function is an S-shaped curve function that can map any real number to the interval (0,1). Subsequently, a threshold of 0.5 is set, and the mapped values ​​are binarized: if the value is greater than or equal to 0.5, it is determined as 1; if the value is less than 0.5, it is determined as 0. This processing method enables the particle swarm optimization algorithm to explore the discrete switching state space in a continuous manner, while obtaining explicit switching instructions during decoding. The binarized 0 and 1 represent two physical states of the power distribution path. 0 indicates that the power distribution path is disconnected, meaning that power cannot be delivered to the corresponding power consumption node through this path; 1 indicates that the power distribution path is connected, meaning that power can be delivered normally through this path. This explicit definition ensures that the optimization results can be directly converted into actual system control commands.

[0091] Based on the above optimal power dispatch strategy, referring to Figure 4 As shown, this application further proposes a specific method for adjusting the output parameters of the generator set and the on / off state of the power distribution path. Specifically, this method includes sending a target output power command to the generator electronic control unit (ECU), which includes a power ramp-up rate limit and a steady-state output power value for the next 5 seconds. After receiving the command, the ECU adjusts the excitation current and fuel flow according to the command, so that the generator output power can smoothly transition to the target value according to the specified ramp-up rate. Simultaneously, this application also includes sending control commands to a solid-state power controller array in the power distribution network, where each solid-state power controller corresponds to a power consumption node or power distribution path. This control command includes the on / off state and a target current limit value. When the solid-state power controller receives the on command, it closes its internal power switch under the condition of not exceeding the target current limit value, thereby completing the reconfiguration of the power transmission path.

[0092] The process involves sending a target output power command to the generator electronic control unit (ECU), including a power ramp-up rate limit and steady-state output power value for the next 5 seconds. This aims to achieve refined and predictive control of the generator's output power. The ECU is typically an embedded system integrating advanced control algorithms (such as PID control and model predictive control), capable of converting received digital commands into analog control signals for the generator's actuators. The power ramp-up rate limit is a key parameter ensuring smooth changes in generator output power; it defines the upper limit of power variation per unit time, effectively preventing system oscillations or shocks caused by sudden power changes. For example, in a turbine-powered propulsion system, the ECU precisely adjusts the excitation current to control the generator's output voltage and reactive power, while simultaneously controlling the mechanical input power of the gas turbine by regulating fuel flow, thereby controlling the generator's active power output. This coordinated adjustment mechanism ensures that the generator's output power can smoothly transition from the current value to the target steady-state value within the predicted 5-second timeframe, following a pre-set smooth curve, thus providing a stable and reliable power supply for the entire system.

[0093] Simultaneously, sending control commands, including on / off status and target current limits, to the solid-state power controller array in the distribution network is the core means to achieve rapid and safe reconfiguration of power transmission paths. A solid-state power controller (SSPC) is an electronic switch based on semiconductor devices (such as MOSFETs and IGBTs). Compared to traditional mechanical relays or circuit breakers, it has significant advantages such as fast response speed, no mechanical wear, long lifespan, frequent operation capability, and easy integration of protection functions. An SSPC array in the distribution network means that multiple SSPCs work collaboratively, each controller responsible for a specific power consumption node or distribution path, thereby achieving refined management of power flow. The on / off status in the control command directly determines whether the corresponding path is connected, while the target current limit serves as an important safety protection measure, ensuring that the current flowing through the path does not exceed a preset safety threshold after the path is connected. Once an overcurrent condition is detected, the SSPC can quickly disconnect the path or limit the current, effectively preventing equipment damage or system failure, thus ensuring system operational safety while completing the reconfiguration of the power transmission path.

[0094] Specifically, in this embodiment, the system operation feedback data includes at least one of the following: real-time voltage, real-time current, real-time power, and generator output frequency at each power-consuming node. Specifically, the real-time voltage of each power-consuming node refers to the potential difference across each node (e.g., propulsion motor, environmental control system, avionics equipment, actuators, etc.) at a specific moment. By connecting high-precision voltage sensors in parallel at the power-consuming nodes, their instantaneous voltage values ​​can be continuously or periodically measured and acquired. These sensors typically have fast response capabilities and high sampling rates to capture rapid voltage changes. Real-time voltage data reflects whether the operating state of the power-consuming node is stable and whether there are abnormal conditions such as undervoltage or overvoltage, which is crucial for assessing the quality of power supply and the health of electrical equipment. Real-time current refers to the charge flow through each power-consuming node at a specific moment. By connecting current sensors (such as Hall effect sensors or shunts) in series on the power supply lines of the power-consuming nodes, their instantaneous current values ​​can be continuously or periodically measured and acquired. Real-time current data directly reflects the actual load condition of the power-consuming node, is the basis for calculating real-time power, and is also an important basis for judging whether the power-consuming node is overloaded or malfunctioning. Real-time power refers to the rate at which electrical energy is actually consumed at each power consumption node at a specific moment. Real-time power can be calculated by multiplying real-time voltage and real-time current (for DC systems) or by considering the power factor (for AC systems). Typically, dedicated power sensors or smart meters can directly measure and output real-time power values. Real-time power data is the most direct indicator for assessing the difference between actual and expected demand at power consumption nodes, and it plays a decisive role in accurately judging system load fluctuations and the effectiveness of dispatching strategies. Generator output frequency refers to the periodic rate of change of the alternating current output by the generator. For turbine-powered propulsion systems, generator output frequency is a key parameter for measuring generator operational stability and power system synchronization. It is obtained by connecting a frequency sensor to the generator output terminal or through a measurement module within the generator's electronic control unit. The stability of the generator output frequency is crucial for the normal operation of the entire power system; deviations may indicate generator overload, speed regulation system malfunctions, or synchronization problems with the grid.

[0095] In practical applications, deviations from system operating states can manifest in various forms, such as persistent systemic deviations or transient, drastic fluctuations. Simply re-predicting may not effectively distinguish and handle these different types of deviations, resulting in inaccurate or untimely regenerated expected power demand when facing complex dynamic load changes, thus affecting the stability and efficiency of overall regulation.

[0096] In this regard, refer to Figure 5 As shown, this application further proposes a method for regenerating the expected electricity demand based on the degree of deviation when feedback data shows a deviation between the actual operating status of the system and the expected dispatch target. Specifically, this includes:

[0097] When system feedback data shows a deviation between the actual operating status and the expected power allocation target, this deviation needs to be precisely quantified. The system collects real-time power values ​​from each power consumption node, reflecting the actual power consumption of each node at the current moment. Simultaneously, based on the previously determined optimal power allocation strategy, there exists a corresponding expected power allocation value, which is the power the system hopes to allocate to each power consumption node. By subtracting the collected real-time power values ​​of each power consumption node from the corresponding expected power allocation value in the optimal power allocation strategy point by point, the real-time power deviation sequence for each power consumption node can be obtained. This deviation sequence intuitively reflects the difference between the actual power and the expected power of each power consumption node at each sampling moment, providing fundamental data for subsequent deviation analysis.

[0098] To gain a more comprehensive understanding of the nature of system deviations, this application further analyzes the real-time power deviation sequence. Specifically, the mean of the real-time power deviation sequence for each power consumption node within the next second is calculated as the steady-state deviation. The mean reflects the average trend or persistent shift of the system deviation within a certain time window, i.e., the systematic prediction deviation. Simultaneously, the variance of this sequence is calculated as the fluctuation deviation. The variance measures the dispersion of the deviation sequence, i.e., the instantaneous fluctuations or instability of the system's operating state, reflecting the prediction fluctuation error. By distinguishing between steady-state deviation and fluctuation deviation, system deviations of different natures can be addressed specifically, avoiding a one-size-fits-all approach.

[0099] To effectively feed the deviation information obtained from the analysis into the prediction of the next cycle, this application introduces a correction mechanism. Specifically, the calculated steady-state deviation and fluctuation deviation are multiplied by a first correction coefficient and a second correction coefficient, respectively. The first correction coefficient is used to compensate for systematic prediction deviations, and its value is positively correlated with the degree of influence of the steady-state deviation on prediction accuracy. This means that when the steady-state deviation is large and has a significant impact on prediction accuracy, the first correction coefficient will increase accordingly to more effectively correct the input. The second correction coefficient is used to suppress prediction fluctuation errors, and its value is positively correlated with the degree of influence of the fluctuation deviation on system stability. This means that when the fluctuation deviation is large and has a significant impact on system stability, the second correction coefficient will increase accordingly to more effectively suppress fluctuations. For example, when there is a persistent underestimation of power demand, the steady-state deviation will be amplified by the first correction coefficient, prompting the next prediction to increase the expected power; when there are severe instantaneous power fluctuations, the fluctuation deviation will be suppressed by the second correction coefficient, so that the prediction results maintain responsiveness while avoiding oversensitivity and unnecessary system oscillations.

[0100] The steady-state and fluctuation deviations, adjusted by correction factors, are then superimposed onto the input vector of the data processing unit for the next cycle. This superposition method allows the data processing unit to fully consider the nature and extent of the current system deviation when making the next prediction, thereby generating a more realistic expected electricity demand. The corrected input vector includes information on the current system operating status, historical electricity consumption records, and intelligently adjusted deviation feedback information. This corrected input vector is then fed into the trained data processing unit. This unit, for example, could be a multi-layer feedforward neural network, which uses its learned complex mapping relationships to recalculate and output the expected electricity demand for each electricity consumption node in the future period based on the new input vector. Through this feedback correction mechanism, the system can dynamically adjust its prediction model to better adapt to changes and uncertainties in the actual operating environment.

[0101] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for dynamic regulation of a turbo-electric propulsion system based on load prediction, characterized in that, Includes the following steps: Collect operational status information and corresponding power consumption records during flight missions; The operational status information and power consumption records are input into a trained data processing unit to generate the expected power demand of each power consumption node in the future period. The trained data processing unit is constructed in the following way: extract no less than 1,000 sets of historical flight mission data from the flight data recorder. Each set of data contains the operational status information at the same moment and the corresponding power consumption records of each power consumption node. The operational status information is used as the input sample and the power consumption records are used as the output label. The backpropagation algorithm is used to iteratively train the multi-layer feedforward neural network. During training, the input samples are propagated forward layer by layer to calculate the network output. The mean square error between the network output and the output label is used as the loss function. The partial derivative of the loss function with respect to the weights of each layer is used as the error signal and propagated backward to update the network weights layer by layer. Repeat the iteration until the loss function value converges to a preset range, resulting in a data processing unit that can output the corresponding power consumption based on the input operating status information. Based on the matching relationship between expected electricity demand and the current output capacity of generators, it is determined whether there are signs of high load fluctuations. This includes: summing the expected electricity demand of each power consumption node in the future period at the same time to obtain the total expected power demand sequence at each time point in the future period; collecting the current maximum output power value of the generator and comparing it with the maximum value in the total expected power demand sequence; calculating the difference between the maximum value and the current maximum output power value of the generator. If the difference is less than 15% of the generator's rated power, it is determined that there are signs of high load fluctuations; at the same time, it is calculated that the power change rate of adjacent time points in the total expected power demand sequence is calculated. If the power change rate of more than three consecutive time points exceeds 20% per second, it is also determined that there are signs of high load fluctuations. If there are signs of high load fluctuations, the optimal power dispatching strategy is determined by iteratively comparing multiple power transmission paths using an optimization algorithm, with the expected power demand as a constraint. Based on the optimal power dispatch strategy, adjust the output parameters of the generator sets and the on / off status of the power distribution channels, and continuously collect system operation feedback data; When feedback data shows that the actual operating status of the system deviates from the expected allocation target, the expected power demand is regenerated based on the degree of deviation, and the process returns to the step of determining the optimal power allocation strategy.

2. The method of claim 1, wherein: Operational status information includes at least one of the following: flight altitude, flight speed, rate of climb, pitch angle, and ambient temperature.

3. The method of claim 1, wherein: The expected power demand of each power consumption node in the future period is represented in the form of power-time series. The specific generation method is as follows: the operating status information of the current moment and the previous 10 consecutive sampling points are combined with the power consumption record to form an input vector, which is input to the trained data processing unit. The output is the power demand value of each power consumption node every 0.5 seconds in the next 5 to 30 seconds, forming a power demand matrix of 60 time points in the future period, including four power consumption nodes: propulsion motor, environmental control system, avionics equipment, and actuation mechanism.

4. The method of claim 1, wherein: The optimization algorithm is a particle swarm optimization algorithm, which iteratively compares multiple options for power transmission paths through the following steps: Initialize a population of 50 particles, each with an eight-dimensional position vector, corresponding to the power allocation ratio of the four main power-consuming nodes and the switching state of the four key power distribution paths. Each particle is randomly initialized with a position and velocity. The current position of each particle is taken as its individual historical best position, and the position with the best fitness value among all particles is taken as the group's historical best position. The fitness function is defined as the sum of the squares of the deviations between the actual power allocated to each power consumption node and the expected power demand, plus a penalty term for the number of switching times of the distribution path. Each particle iterates according to its own historical best position and the group's historical best position, following the velocity update formula and the position update formula. The velocity update formula includes inertia weight, cognitive learning factor, and social learning factor. After each iteration, the fitness value of each particle is recalculated, and the individual historical best position and the group historical best position are updated. Repeat the iteration until the preset number of iterations of 200 is reached or the group's historical best position remains unchanged for 30 consecutive iterations. Then, decode the eight-dimensional vector corresponding to the final group's historical best position into the power allocation ratio of each power consumption node and the switching state of the power distribution path, which will serve as the optimal power dispatching strategy.

5. The dynamic control method for a turbine-powered propulsion system based on load prediction according to claim 4, characterized in that: In the particle swarm optimization algorithm, the position vector encoding method for each particle is as follows: The first four dimensions of the position vector correspond to the power distribution ratio of the propulsion motor, environmental control system, avionics equipment, and actuator, respectively. The values ​​of the first four dimensions range from 0 to 1, and the sum of the four is 1. This is achieved by performing softmax normalization on the original position values. The last four dimensions of the position vector correspond to the switching states of the first power distribution path connecting the generator and the propulsion motor, the second power distribution path connecting the generator and the environmental control system, the third power distribution path connecting the generator and the avionics equipment, and the fourth power distribution path connecting the generator and the actuation mechanism. The values ​​of the last four dimensions are mapped to the interval between 0 and 1 by the sigmoid function, and then binarized to 0 or 1 with a threshold of 0.5, where 0 represents disconnection and 1 represents connection.

6. The dynamic control method for a turbine-powered propulsion system based on load prediction according to claim 1, characterized in that: Adjusting the generator set output parameters and the on / off status of the power distribution path, specifically including: The generator electronic control unit sends a target output power command, which includes the power ramp-up rate limit and steady-state output power value for the next 5 seconds. The generator electronic control unit adjusts the excitation current and fuel flow according to the command, so that the generator output power smoothly transitions to the target value at the specified ramp-up rate. Simultaneously, control commands are sent to the solid-state power controller array in the power distribution network. Each solid-state power controller corresponds to a power consumption node or power distribution path. The control commands include on / off status and target current limit value. After receiving the on command, the solid-state power controller closes its internal power switch under the condition of not exceeding the target current limit value, thus completing the reconfiguration of the power transmission path.

7. The dynamic control method for a turbine-powered propulsion system based on load prediction according to claim 1, characterized in that: The system operation feedback data includes at least one of the following: real-time voltage, real-time current, real-time power, and generator output frequency at each power consumption node.

8. The dynamic control method for a turbine-powered propulsion system based on load prediction according to claim 1, characterized in that: When feedback data shows a deviation between the actual operating status of the system and the expected allocation target, the expected electricity demand is regenerated based on the degree of deviation, specifically including: The real-time power values ​​of each power consumption node are collected and subtracted point by point from the expected power allocation value in the optimal power dispatch strategy to obtain the real-time power deviation sequence of each power consumption node. The mean of the real-time power deviation sequence of each power consumption node within the next second is calculated as the steady-state deviation, and the variance of the sequence is calculated as the fluctuation deviation. The steady-state deviation and the fluctuation deviation are multiplied by the first correction coefficient and the second correction coefficient, respectively, and then superimposed on the input vector of the data processing unit in the next cycle. The first correction coefficient is used to compensate for systematic prediction deviation, and its value is positively correlated with the degree of influence of steady-state deviation on prediction accuracy. The second correction coefficient is used to suppress prediction fluctuation error, and its value is positively correlated with the degree of influence of fluctuation deviation on system stability. The corrected input vector is input to the trained data processing unit to regenerate the expected power demand of each power consumption node in the future period.

Images