Self-learning correction power prediction model online updating method and system

Abstract

The invention discloses a self-learning correction power prediction model online updating method and system, and belongs to the technical field of artificial intelligence and data-driven modeling, and the method comprises the steps: obtaining an input data set related to target power, constructing a preliminary prediction model for prediction, and generating a preliminary power prediction value; obtaining an actual measurement power value corresponding to the preliminary power prediction value in the same period, comparing the actual measurement power value with the preliminary power prediction value, calculating a prediction deviation value, extracting dynamic characteristics, generating a health state fingerprint, matching the health state fingerprint with a preset deviation mode library, generating a matching result, triggering a correction strategy, and adjusting the subsequently generated preliminary power prediction value. A modified power prediction value is generated. According to the method, the technical means of extracting the prediction deviation dynamic features to generate the health state fingerprints and matching the health state fingerprints with the deviation pattern library to determine the correction strategy is adopted, online diagnosis and target-clear self-adaptive correction can be carried out on the prediction model, and therefore the long-term accuracy and the self-adaptive capacity of prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and data-driven modeling, in particular to a self-learning correction power prediction model online updating method and system. BACKGROUND

[0002] Power prediction is a key technology in the fields of energy management, power grid dispatching and industrial production, which aims to estimate future power output or consumption based on historical operation data of equipment, real-time operating conditions and related environmental factors. With the development of artificial intelligence technology, using machine learning models such as neural networks for power prediction has become mainstream. Such models can establish an accurate mapping from multi-dimensional input to power output by learning the complex nonlinear relationships in historical data, thereby providing data-driven prediction capabilities.

[0003] In the related art, the Chinese invention with publication number CN120433304B discloses a photovoltaic energy storage and direct flexible system power scheduling method based on market model prediction response, which includes determining the composition of the photovoltaic energy storage and direct flexible system, collecting operation state data and external environment data; constructing a hybrid prediction model to predict market prices and uncertainty intervals of market prices within a scheduling period; establishing a power scheduling optimization model to obtain the optimal power scheduling strategy; according to the optimal power scheduling strategy, control instructions are issued to each control unit of the photovoltaic energy storage and direct flexible system, and collaborative adjustment is performed through a bidirectional coupling feedback mechanism of voltage and power; real-time monitoring of market prices and system operation state, when the market price deviation exceeds the price deviation threshold, triggering online correction of the scheduling strategy.

[0004] For the related technology in the above, the inventor believes that the above scheme relies on a single and coarse-grained index of "market price deviation exceeding the price deviation threshold" to trigger the scheduling strategy correction, which has insufficient prediction deviation analysis granularity; only describes "triggering online correction of the scheduling strategy", but does not elaborate on the specific correction method, strategy type and how to adjust according to the deviation type; and does not explicitly propose a mechanism for performance evaluation, diagnosis or self-evolution of the prediction model itself according to the correction effect. SUMMARY

[0005] To solve the above problems, the present application provides a self-learning correction power prediction model online updating method and system, which uses the technical means of extracting prediction deviation dynamic features to generate health status fingerprints, and matching with deviation pattern library to determine correction strategy, which can perform online diagnosis and target-oriented adaptive correction on the prediction model, thereby improving the long-term accuracy of prediction and the autonomous adaptation ability of the system.

[0006] The above object can be achieved by the following scheme:

[0007] The application discloses an online updating method and system for a self-learning corrected power prediction model, which comprises the following steps: acquiring an input data set of real-time operation conditions and environmental influence factors related to target power; constructing a preliminary prediction model; predicting the input data set based on the preliminary prediction model to generate a preliminary power prediction value; acquiring an actual measured power value corresponding to the preliminary power prediction value at the same period and comparing the actual measured power value with the preliminary power prediction value to calculate a prediction deviation value; extracting a dynamic feature of the prediction deviation value to generate a health state fingerprint; matching the health state fingerprint with a preset deviation mode library to determine a correction strategy and generate a matching result; triggering the corresponding correction strategy based on the matching result and adjusting the preliminary power prediction value generated subsequently to generate a corrected power prediction value.

[0008] Optionally, the generating of the preliminary power prediction value comprises the following steps: collecting real-time operation conditions of equipment operation, which comprises operation load, start-stop state, equipment operation duration and operation mode switching information, and simultaneously acquiring environmental influence factors, which comprises temperature, humidity, wind speed, atmospheric pressure and solar radiation intensity data, to form an input data set; constructing a preliminary prediction model; and performing dynamic feature extraction and trend fitting on the input data set based on the preliminary prediction model to generate a preliminary power prediction value.

[0009] Optionally, the constructing of the preliminary prediction model comprises the following steps: constructing a prediction structure with dynamic response capability based on historical operation data of a target equipment and the input data set; performing multi-scale feature fusion on the input data set based on the prediction structure to capture the internal correlation between short-term fluctuations and long-term trends and establish a mapping relationship from the real-time operation conditions and the environmental influence factors to power output; performing adaptive weight adjustment on the mapping relationship, dynamically adjusting the contribution degree according to the real-time strength of each element in the input data set on power, and obtaining an optimized feature expression; and performing parameter tuning of the prediction structure based on the optimized feature expression through an iterative optimization process to obtain a preliminary prediction model.

[0010] Optionally, the generating of the health state fingerprint comprises the following steps: performing feature extraction based on the prediction deviation value to obtain a historical prediction deviation value sequence; performing trend analysis on the historical prediction deviation value sequence to generate a deviation change trend slope; performing statistical analysis on the historical prediction deviation value sequence to generate a deviation duration positive-negative nature and a fluctuation frequency; and combining the deviation change trend slope, the deviation duration positive-negative nature and the fluctuation frequency to generate a health state fingerprint.

[0011] Optionally, the method further comprises: obtaining a device running state log, correlatively analyzing the health state fingerprint with the device running state log, and identifying an abnormal fingerprint pattern related to running mode switching and load mutation; based on the abnormal fingerprint pattern, generating a device health warning signal, and triggering local parameter fine-tuning of the preliminary prediction model.

[0012] Optionally, the generating a matching result comprises: reading multiple state patterns from the deviation pattern library, wherein the state pattern is jointly represented by the deviation change trend slope, the deviation duration positive and negative nature, and the fluctuation frequency; calculating the similarity of the health state fingerprint with each state pattern to generate a similarity score set; based on the similarity score set, screening the state pattern to obtain a matching result.

[0013] Optionally, the generating a corrected power prediction value comprises: determining the type of the corresponding correction strategy according to the matching result, and applying dynamic compensation to the preliminary power prediction value, wherein the dynamic compensation comprises trend compensation, bias compensation, or fluctuation suppression; after implementing the dynamic compensation, performing a reasonableness check on the compensated prediction value, and if the check passes, outputting a corrected power prediction value, otherwise, reverting to an uncompensated state and retriggering the correction strategy selection step; outputting the corrected power prediction value as a final prediction result, and synchronously updating the deviation pattern library.

[0014] Optionally, the method further comprises: obtaining a subsequent prediction deviation value of the corrected power prediction output and the actual measured power value; based on the subsequent prediction deviation value, updating the health state fingerprint, and monitoring an indicative performance; when the indicative performance continuously decreases, triggering a comprehensive recalibration of the preliminary prediction model.

[0015] Optionally, the triggering a comprehensive recalibration of the preliminary prediction model comprises: during the recalibration process, integrating the historical running data, the real-time operation condition, and the environmental influence factor to construct a global training sample set; based on the global training sample set, using a preset incremental learning algorithm to reconstruct the structure and reset the parameters of the preliminary prediction model, to generate a new generation prediction model; deploying the new generation prediction model, and resetting the health state fingerprint and the deviation pattern library.

[0016] Based on the same inventive concept, the application also provides a self-learning correction power prediction model online updating system, which comprises: a power prediction module, which is used for acquiring an input data set of real-time operating conditions and environmental influence factors related to target power, constructing a preliminary prediction model, predicting the input data set based on the preliminary prediction model, and generating a preliminary power prediction value; a deviation analysis module, which is used for acquiring and comparing actual measured power values corresponding to the preliminary power prediction value at the same period, calculating a prediction deviation value, extracting dynamic characteristics of the prediction deviation value, and generating a health state fingerprint; a strategy matching module, which is used for matching the health state fingerprint with a preset deviation mode library, determining a correction strategy, and generating a matching result; and a prediction correction module, which is used for triggering the corresponding correction strategy based on the matching result, adjusting the preliminary power prediction value generated subsequently, and generating a corrected power prediction value.

[0017] Compared with the prior art, the application has the following advantages:

[0018] 1. The application realizes deep diagnosis and targeted correction of model performance by constructing a dynamic characteristic fingerprint of prediction deviation and matching it with a preset deviation mode library. The method can intelligently distinguish the root cause of deviation, such as trend drift, systematic bias or random fluctuation, and trigger the most appropriate dynamic compensation measures, reducing the blindness of traditional error correction methods and improving the accuracy and efficiency of correction.

[0019] 2. The application establishes a multi-level self-adaptive maintenance mechanism, realizing automatic management of the whole life cycle of the prediction model. By introducing local parameter fine-tuning associated with device operation logs and triggering comprehensive recalibration when the correction effect continuously decreases, different levels of coping strategies can be taken according to the severity of the problem, from lightweight adjustment for rapid response to fundamental model reconstruction, enhancing the robustness and long-term reliability of the prediction system under complex dynamic working conditions.

[0020] 3. The application brings the performance of the corrected system into a new round of monitoring closed loop and updates the deviation mode library synchronously, giving the whole system the ability of continuous learning and evolution. The method not only corrects the current prediction result, but also learns from successful correction experience, constantly enriches and optimizes its diagnosis knowledge base and strategy library, so that the intelligence level and correction ability of the system are improved simultaneously with the growth of running time.

[0021] Other features and advantages of the application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the application. The objects and other advantages of the application will be realized and attained by the structure particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0023] Figure 1 is a flowchart of an online updating method of a self-learning and modified power prediction model according to an embodiment of the present application.

[0024] Figure 2 is a dynamic characteristic analysis diagram of a prediction deviation according to an embodiment of the present application.

[0025] Figure 3 is a schematic diagram of performance monitoring and comprehensive recalibration triggering according to an embodiment of the present application.

[0026] Figure 4 is a structural schematic diagram of an online updating system of a self-learning and modified power prediction model according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort belong to the protection scope of the present application.

[0028] With reference to Figure 1 , one embodiment of the present application proposes an online updating method of a self-learning and modified power prediction model, which adopts a technical means of extracting dynamic characteristics of a prediction deviation to generate a health state fingerprint, and matching the deviation mode library to determine a correction strategy, so as to perform online diagnosis and target-oriented adaptive correction on the prediction model, thereby improving the long-term accuracy of prediction and the autonomous adaptation ability of the system.

[0029] The method according to the embodiment specifically includes:

[0030] obtaining an input data set of real-time operating conditions and environmental influence factors related to target power, constructing a preliminary prediction model, predicting the input data set based on the preliminary prediction model, and generating a preliminary power prediction value;

[0031] Optionally, the generating of the preliminary power prediction value includes:

[0032] The system collects real-time operating conditions of the equipment, including operating load, start / stop status, equipment runtime, and operating mode switching information. It also acquires environmental influencing factors, including data on temperature, humidity, wind speed, atmospheric pressure, and solar radiation intensity, to form an input dataset.

[0033] A preliminary prediction model is constructed, and dynamic feature extraction and trend fitting are performed on the input dataset based on the preliminary prediction model to generate preliminary power prediction values.

[0034] Specifically, the system first collects real-time operating conditions of the target equipment, including various indicators directly reflecting its working status, such as operating load (the percentage of current power output or load handled), start / stop status (a Boolean value indicating whether the equipment is on or off), equipment runtime (the cumulative running time since the last start / stop or maintenance), and operating mode switching information to distinguish between different operating states, such as energy-saving mode or high-performance mode. Simultaneously, it acquires external environmental factors that significantly affect the equipment's power performance, covering temperature, humidity, wind speed, atmospheric pressure, and, in specific application scenarios such as solar radiation intensity in photovoltaic power generation. The collected real-time operating conditions and environmental factors are combined into a multi-dimensional feature vector, which constitutes the input dataset. A pre-built preliminary prediction model is then used to analyze the input dataset. This preliminary prediction model is capable of processing time-series data. Upon receiving the input, it first performs dynamic feature extraction, aiming to identify and quantify dynamic features that are indicative of power changes from the time-series information, such as the rate of change, periodic fluctuations, and short-term disturbances, and then performs unified data standardization processing. Next, trend fitting is performed, which involves capturing and quantifying the long-term trends, seasonal variations, and periodic patterns of the power data based on the extracted dynamic features. This is typically achieved through multiple regression analysis, linear time series models such as exponential smoothing models, or more complex machine learning models such as long short-term memory networks. The model uses dynamic features as input variables; if a linear relationship is established between the dynamic features and the target power value, it can predict future power trends under current operating conditions and environmental influences, thus aiding in the calculation of preliminary power predictions. ,have:

[0035] ;

[0036] in, Based on the baseline power value; For the extracted first One dynamic feature; For the first Features The slope parameter of power is determined by least square fitting of training data, i.e. finding the parameter combination that minimizes the sum of square error between predicted and actual values. The final generated preliminary power prediction value will be used as the baseline input for the subsequent bias analysis and model correction stage.

[0037] Exemplarily, the process of generating preliminary power prediction value starts from data collection. Take an industrial gas turbine as an example, its real-time operating conditions are collected every minute, including the running load percentage, which is 75% at this moment; the start-stop state is “running”; the equipment has been running for 120 hours; the operation mode is “standard mode”. At the same time, environmental impact factors at the same time are obtained, such as environmental temperature of 25 degrees Celsius, humidity of 60%, wind speed of 5 meters per second, atmospheric pressure of 101 kilopascals, and solar radiation intensity data is not required here. These numerical values are combined into a multi-dimensional feature vector, and the time stamp is recorded to form a data point in the input data set. These data points are arranged in time sequence to form a complete input data set for model processing. Then the input data set is input into the pre-trained preliminary prediction model. The preliminary prediction model is a long short-term memory network (LSTM) model based on deep learning. After receiving the input, it will first extract dynamic features from the time series data through its internal gating mechanism, such as identifying the load change rate, the running trend before mode switching, etc. Then based on these dynamic features, trend fitting is completed to infer the expected power output of the gas turbine in the next 15 minutes. For example, output a preliminary power prediction value for the next 15 minutes, the value is 88.5 megawatts. The preliminary power prediction value will be used as the baseline for subsequent processing. By comprehensively integrating diverse real-time operating conditions and environmental impact factors, a preliminary power prediction value reflecting the complex working state of the equipment and environmental impact can be generated, improving the accuracy and comprehensiveness of the initial prediction.

[0038] Optionally, the constructing a preliminary prediction model comprises:

[0039] constructing a prediction structure with dynamic response capability based on historical operation data of the target equipment and the input data set;

[0040] performing multi-scale feature fusion on the input data set based on the prediction structure, capturing the internal relationship between short-term fluctuations and long-term trends, and establishing a mapping relationship from the real-time operating conditions and the environmental impact factors to the power output;

[0041] performing adaptive weight adjustment on the mapping relationship, dynamically adjusting the contribution degree according to the real-time strength of each element in the input data set on the power, and obtaining an optimized feature expression;

[0042] obtaining the preliminary prediction model through an iterative optimization process based on the optimized feature expression.

[0043] Specifically, first, a prediction structure with dynamic response capability is constructed based on the historical operation data of the target device and the structural characteristics of the current input data set. This prediction structure usually selects a model that can effectively process time series information, such as long short-term memory network (LSTM) or gated recurrent unit (GRU), because the internal gating mechanism can learn and remember the time dependence in the data, thereby making dynamic responses to changes in input. After the prediction structure is established, multi-scale feature fusion is performed on the input data set. This process aims to capture both short-term fluctuations and long-term trends in power data. For example, the model can be designed to process two inputs in parallel. One input is the data of the last few minutes or hours, which is used to capture the short-term fluctuations and immediate response of the device operation; the other input is the data of the past few days or weeks, which is used to learn seasonal, workday periodicity and other long-term trends. Through the fusion layer, the features extracted from different time scales are integrated to obtain the mapping relationship from real-time operating conditions and environmental influencing factors to power output. Then, in order to enable the model to dynamically focus on the most important influencing factors, adaptive weight adjustment is performed on the mapping relationship. This step is usually implemented by introducing an attention mechanism. The attention mechanism dynamically calculates a contribution weight for each input feature or each time step feature according to the real-time strength of each element in the input data set on the power. For calculating the contribution weight of the m-th input feature ,

[0044] ;

[0045] where is a da×dh feature projection matrix, is a da×dq query projection matrix, and these two matrices are learnable parameters in the attention mechanism; is a dh×1 vector representation of the m-th feature source; is a dq×1 query vector, usually from the decoder state, the last hidden state or a learnable parameter; is a da×1 attention weight vector, which is a learnable parameter; is the hyperbolic tangent function; ​is the total number of input features. da, dh, dq represent the number of rows and columns in different matrices. The higher the contribution weight of a feature, the greater its influence on the current prediction time. By applying these adaptive weights to the multi-scale fused features, a weighted and more refined optimized feature representation can be generated. Finally, the optimized feature representation is input into the top layer of the prediction structure for the final power value prediction, and the prediction result is compared with the actual power value in the historical data. Then, using optimization algorithms such as gradient descent, the error is backpropagated and all learnable parameters inside the prediction structure are iteratively updated, including the weights of the LSTM units, the parameters of the multi-scale fusion layer, and the parameters of the attention mechanism. This iterative optimization process continues until the model's performance on the validation set reaches the preset convergence criteria. After sufficient parameter tuning, the resulting model is the final preliminary prediction model.

[0046] For example, when building a preliminary prediction model, first, based on the historical operation data of the gas turbine and the real-time input data set, a multi-scale LSTM prediction structure with dynamic response capability is constructed. Let the input feature vector dimension be dh=64, including real-time and environmental data such as load, temperature, etc., the query vector dimension dq=64, and the attention hidden dimension da=128. The model processes two time scale inputs in parallel: past 1 minute level data and past 7 days hour level average data, which are extracted by LSTM and then concatenated in the fusion layer to form a multi-dimensional to power mapping relationship. To dynamically adjust the importance of features, an additive attention mechanism is introduced after the fusion layer, which includes feature projection matrix, query projection matrix and attention weight vector. Taking three features h1 load, h2 temperature, h3 wind speed as an example, the contribution weight is calculated: 0.652, 0.217, 0.131, with three significant digits. This indicates that the load feature has the highest proportion in the current prediction. Finally, based on the weighted fused features, a preliminary power prediction value is generated, and an optimizer is used to iteratively optimize all parameters on the historical data training set until convergence, thereby obtaining a preliminary prediction model with adaptive attention adjustment capability.

[0047] An actual measured power value corresponding to the same period as the preliminary power prediction value is obtained and compared, a prediction deviation value is calculated, dynamic features of the prediction deviation value are extracted, and a health status fingerprint is generated;

[0048] Optionally, the generating a health status fingerprint comprises:

[0049] Based on the prediction deviation value, feature extraction is performed to obtain a historical prediction deviation value sequence;

[0050] Trend analysis is performed on the historical prediction deviation value sequence to generate a deviation change trend slope.

[0051] Statistical analysis is performed on the historical prediction deviation value sequence to generate the duration of the deviation being positive or negative and the fluctuation frequency.

[0052] A health status fingerprint is generated by combining the slope of the deviation change trend, the duration of the deviation being positive or negative, and the fluctuation frequency.

[0053] Specifically, the process first extracts features based on a series of continuous prediction deviation values, where the prediction deviation is the difference between the initial power prediction and the actual measured power value during the same period. A fixed-length time window, such as 60 minutes, is maintained, and all prediction deviation values ​​within this time window are collected, resulting in a time series, i.e., the historical prediction deviation value series. Next, trend analysis is performed on this historical prediction deviation value series to generate the slope of the deviation change trend. This step aims to quantify whether the overall evolution of the prediction deviation over a period of time is increasing, decreasing, or remaining stable. The calculation of the deviation change trend slope is then performed. ,have:

[0054] ;

[0055] in, At a certain point in time The prediction deviation value at that location, It is the total number of data points in the historical prediction deviation value sequence; It is the average value at a given time point; This is the average of the prediction bias values. A positive slope indicates that the prediction bias is gradually increasing, a negative slope indicates that it is decreasing, and a slope close to zero indicates that the bias is stabilizing. For example... Figure 2As shown, the black dotted line in the figure depicts the actual data points of the difference between the preliminary power prediction value and the actual measured power value, i.e. the prediction bias value, over time; the black dashed line represents a trend fitting line obtained by linear regression fitting on the historical prediction bias value sequence. Meanwhile, statistical analysis is performed on the historical prediction bias value sequence to generate bias persistence and fluctuation frequency. Bias persistence is used to describe whether the prediction bias is continuously positive or negative, which quantifies the degree of systematic overestimation or underestimation of the model; it is obtained by counting the maximum length of time or the number of data points in which the sign of the prediction bias value remains unchanged within the current time window. Fluctuation frequency describes the frequency of changes in the sign of the prediction bias value. It can be obtained by counting the number of times the historical prediction bias value sequence crosses the zero axis within the time window, and then dividing by the total length of the time window, for example, if the bias value changes from positive to negative or from negative to positive W times within a 60-minute time window, then the fluctuation frequency is W times / 60 minutes. Finally, the three dynamic characteristics of the bias change trend slope, bias persistence and fluctuation frequency obtained through the above analysis are combined to form a multi-dimensional feature vector. This feature vector can comprehensively characterize the current performance and bias characteristics of the prediction model from the aspects of trend, bias and stability. The combined vector is defined as the health status fingerprint.

[0056] Exemplarily, the process of generating the health status fingerprint is based on the bias data of the preliminary power prediction value and the actual measured value of the gas turbine. Assuming that within the past 60 minutes, the bias sequence of the preliminary power prediction value and the actual measured power value is as follows: 5-minute interval sampling values, in which the bias sequence corresponding to time points 1 to 12 is MW, the time sequence is The total number of data points is 12. The bias change trend slope is calculated as Megawatts per minute, the result is retained to three decimal places, indicating that the prediction deviation presents a gradually increasing trend over time. Then calculate the deviation persistence positive and negative. In the above deviation sequence, all deviation values are positive. Therefore, the deviation persistence positive and negative can be expressed as "the duration of consecutive positive values is 60 minutes", or "the number of consecutive positive data points is 12", which indicates that the model has a systematic tendency to underestimate. Subsequently, the system calculates the fluctuation frequency. In this sequence, since all deviation values are positive, there is no case of crossing the zero axis, so the fluctuation frequency is 0 times per 60 minutes, indicating that the error is very stable and there is no frequent alternation of positive and negative values. Finally, the three features of the calculated deviation change trend slope, deviation persistence positive and negative "persistent positive", and fluctuation frequency 0 times per 60 minutes are combined to form a health status fingerprint vector, such as [0.275, "persistent positive", 0]. This fingerprint comprehensively characterizes the deviation characteristics of the current preliminary prediction model. Through comprehensive and in-depth dynamic analysis of the deviation between the preliminary power prediction value and the actual measured value, the performance state of the model can be quantified, and the perception and management ability of the prediction performance is improved.

[0057] Optionally, the method further comprises:

[0058] Obtaining a device running state log, correlating and analyzing the health status fingerprint with the device running state log, and identifying an abnormal fingerprint pattern related to running mode switching and load mutation;

[0059] Based on the abnormal fingerprint pattern, a device health warning signal is generated, and a local parameter fine-tuning for the preliminary prediction model is triggered.

[0060] Specifically, first is to actively acquire the device operation state logs, which are independent data sources from real-time operating conditions, usually recorded by the device's control system or supervisory control and data acquisition system, and the content is discrete events with precise timestamps, such as recording the running mode switching at a certain time, or the sharp change of load at a certain time point, i.e. load mutation. After obtaining the logs, the continuously generated health status fingerprints are associated and analyzed with the device operation state logs. The core of this analysis is the alignment in time and the exploration of causal relationship. Each key event extracted from the logs, such as a running mode switching, will be taken as a time anchor point. Then the health status fingerprint sequence generated within a small time window after the event will be reviewed. Through statistical learning of historical events and their subsequent fingerprints, strong correlation between specific events and specific fingerprint patterns can be identified. For example, it may be found that after each load mutation, the health status fingerprint generally shows a sharp increase in the absolute value of the deviation change trend slope and a significant increase in fluctuation frequency. This fingerprint pattern closely coupled with a specific physical event is identified as an abnormal fingerprint pattern. Once in the online monitoring process, the newly generated health status fingerprint matches the known abnormal fingerprint pattern, two parallel operations will be performed. First, generate a device health warning signal. This signal can be an alarm sent to the operation and maintenance personnel, or a highlighted sign on the monitoring interface, whose purpose is to prompt from the business level that the current prediction deviation increase may be caused by specific device operation, not the universal decline of the model itself. Second, the matching result will trigger the local parameter fine-tuning of the preliminary prediction model. Local parameter fine-tuning is a lightweight and fast response updating mechanism that is different from full-scale recalibration. It does not reset all parameters of the model, but according to the event type associated with the abnormal fingerprint pattern, it makes a small adjustment to the most relevant part of the preliminary prediction model. If it is identified that the abnormality is caused by load mutation, the output layer of the model will be trained for 3 to 5 iterations using the latest device operation data and actual power values related to the mutation event, so that it can quickly adapt to the new working state of the device.

[0061] Exemplarily, the method is directed to a gas turbine power prediction system. After generating the health state fingerprint, the method further combines with the device operation log to perform intelligent early warning and model fine-tuning. For example, at 3:05 pm on a certain day, the device control system log records a "load mutation event, the load rises sharply from 70% to 95%". Within the next 10 minutes, a new health state fingerprint [1.5, "mixed positive and negative", 5 times / 10 minutes] is collected and generated. By comparing historical data, it is found that in the past, when the load mutates, the health state fingerprint usually shows that the absolute value of the deviation change trend slope significantly increases, for example, jumps from the usual 0.1 to 1.0 or above; at the same time, the fluctuation frequency also significantly increases from the usual 1-2 times / 10 minutes to 5 times / 10 minutes or above, and the "mixed positive and negative" deviation may appear alternately. The generated fingerprint [1.5, "mixed positive and negative", 5 times / 10 minutes] is highly matched with this known "load mutation" abnormal fingerprint pattern. Once the match is identified, the system immediately performs two operations. First, a device health warning signal is generated. It will send an alert message to the operation and maintenance personnel "gas turbine power prediction deviation anomaly, which may be related to the load mutation at 3:05 pm, please pay attention to the current prediction performance of the model". Second, local parameter fine-tuning is triggered for the preliminary prediction model. Recognizing that the anomaly is related to the load change, the last 5 minutes of high-frequency real data containing the load mutation event are used to perform 3 times of gradient update on the output layer of the preliminary prediction model, so that it quickly adapts to the new high-load working state of the device. By associating the dynamic health state fingerprint with the accurate device operation state log, the abnormal prediction pattern related to a specific physical event can be identified in time, and the real-time adaptability and robustness of the prediction system are improved.

[0062] The health state fingerprint is matched with a preset deviation mode library to determine a correction strategy, and a matching result is generated.

[0063] Optionally, the generating a matching result comprises:

[0064] A plurality of state modes are read from the deviation mode library, wherein the state mode is jointly represented by the deviation change trend slope, the deviation continuous positive and negative nature, and the fluctuation frequency.

[0065] The similarity between the health state fingerprint and each state mode is calculated to generate a similarity score set.

[0066] Based on the similarity score set, the state mode is screened to obtain a matching result.

[0067] Specifically, first, a plurality of state patterns are read from a preset deviation pattern library. The deviation pattern library is a database that is constructed in advance and can be dynamically updated, and a large number of state patterns are stored in the database. Each state pattern is a feature vector with the same structure as the health state fingerprint, and is jointly represented by three dimensions of deviation change trend slope, deviation persistence positive and negative, and fluctuation frequency. Each state pattern represents a known and typical prediction deviation behavior, such as “persistent positive bias”, “high-frequency random fluctuation” or “gradual negative drift”, and each state pattern is pre-associated with one or more recommended correction strategies. Next, the similarity between the currently generated health state fingerprint and each state pattern in the deviation pattern library is calculated. This is a measurement process in a vector space. In order to eliminate the differences in dimension and numerical range between different dimensional features, it is necessary to first normalize the components of the health state fingerprint and all state patterns, i.e., the deviation change trend slope, the deviation persistence positive and negative, and the fluctuation frequency. After normalization, the Gaussian radial basis function kernel can be used to calculate the similarity score. For calculating the similarity score between the health state fingerprint and the i-th state pattern in the deviation pattern library, there is:

[0068]

[0069] wherein,

[0070] is the normalized health state fingerprint vector and the i-th state pattern vector, respectively; represents the square of the Euclidean distance between the two normalized vectors; is an adjustable parameter used to control the sensitivity of the similarity function to the distance. By repeating this calculation for each state pattern in the deviation pattern library, a set containing all similarity scores, i.e., the similarity score set, is obtained. Finally, based on the similarity score set, the state pattern is screened to obtain the final matching result. The most direct screening method is to select the state pattern with the highest similarity score as the best match. In practice, a similarity threshold is usually set, and only when the highest similarity score exceeds the similarity threshold, is the matching considered successful. This matching successful state pattern is the final matching result. If all similarity scores are lower than the similarity threshold, it may indicate that the current deviation is a new unknown pattern, which can be marked and trigger the subsequent learning and updating mechanism.

[0069] Exemplarily, after obtaining a new health state fingerprint, such as a health state fingerprint vector [0.2748, "persistent positive", 0], the gas turbine power prediction system matches it with a preset deviation mode library. The deviation mode library stores multiple normalized state modes, each of which is associated with a correction strategy, such as mode one [0.2, "persistent positive", 0.1], associated with a correction strategy of "positive bias compensation". Mode two [-0.1, "persistent positive", 0.2], associated with a correction strategy of "negative bias compensation". Mode three [0.05, "mixed positive and negative", 0.8], associated with a correction strategy of "high-frequency fluctuation suppression". In order to calculate the similarity between the health state fingerprint and these modes, it is necessary to normalize each component in the fingerprint and the mode first. For example, normalize the deviation trend slope and fluctuation frequency to 0 to 1, and convert the deviation persistent positive and negative into numerical features using hot encoding. Assuming that after normalization, the current health state fingerprint is [0.8, 1, 0.0], where 1 represents persistent positive, and mode one is [0.7, 1, 0.05]. Next, the Gaussian radial basis function kernel is used to calculate the similarity, , and the result is rounded to three significant digits. Repeat this calculation for mode two and mode three, and assume that the results are , The similarity score set is obtained . Finally, based on the score set, the highest similarity score is selected . If the preset similarity threshold is 0.5, since 0.535 is greater than 0.5, mode one is selected as the best matching result. Therefore, the matching result indicates that the "positive bias compensation" correction strategy should be triggered. By normalizing the matching between the real-time generated health state fingerprint and the preset and instructive deviation mode library, the specific type of the current prediction deviation can be quickly and accurately diagnosed.

[0071] Based on the matching result, the corresponding correction strategy is triggered, and the subsequently generated preliminary power prediction value is adjusted to generate a corrected power prediction value.

[0072] Optionally, the generation of the corrected power prediction value includes:

[0073] According to the matching result, the type of the corresponding correction strategy is determined, and a dynamic compensation is applied to the preliminary power prediction value, wherein the dynamic compensation includes trend compensation, bias compensation or fluctuation suppression;

[0074] After implementing the dynamic compensation, the compensated prediction value is reasonably checked. If the check is passed, the output is the corrected power prediction value, otherwise, it is rolled back to the uncompensated state and the correction strategy selection step is retriggered.

[0075] The modified power prediction value is output as the final prediction result, and the updating of the deviation mode library is performed synchronously.

[0076] Specifically, this process begins with determining the type of correction strategy that should be performed according to the matching results generated in the previous stage. The state mode in the deviation mode library pointed to by the matching results is pre-associated with a specific correction strategy. If the matching state mode is "progressive negative drift", the deviation change trend slope is negative, and the deviation is continuously negative, the trend compensation strategy is triggered; if the matching state mode is "persistent positive bias", the deviation is continuously positive and the fluctuation is small, the bias compensation strategy is triggered; if the matching state mode is "high-frequency random fluctuation", the fluctuation frequency is high, the fluctuation suppression strategy is triggered. After determining the correction strategy, the corresponding dynamic compensation is applied to the newly generated preliminary power prediction value. Trend compensation may be to add a correction amount that increases linearly over time to the preliminary prediction value; bias compensation is to directly add or subtract a fixed correction amount to the preliminary prediction value; and fluctuation suppression may be to smooth the output by moving average or low-pass filtering on the prediction value sequence. After implementing dynamic compensation, the compensated prediction value is checked for reasonableness. This is a key quality control step, aiming to prevent the prediction result from deviating from the physical reality due to inappropriate compensation. The checking rules can include checking whether the modified power prediction value exceeds the maximum rated power of the device or is lower than the minimum operating power of the device, or checking whether its change rate exceeds the maximum power ramping rate possible in the physical device. If the check passes, it means that the modified prediction value is reasonable in terms of physics and operation, and the value will be officially output as the modified power prediction value. If the check fails, it means that the correction strategy application fails, and it will be rolled back to the uncompensated state, i.e. giving up this compensation operation, and it may re-trigger the correction strategy selection step, such as selecting the state mode with the second highest similarity score for trial, or recording this failure as a basis for future updating of the deviation mode library. The modified power prediction value that passes the final check will be output as the final prediction result to the user or downstream application system. At the same time, an important synchronous operation is to update the deviation mode library. If the current correction significantly improves the prediction accuracy, for example, if the average absolute error of the original preliminary power prediction value is 5%, and after correction by the correction strategy, the average absolute error is reduced to 2%, then this reduction can be judged as a significant improvement in prediction accuracy. This time's health status fingerprint, selected correction strategy, and correction effect can be recorded, and if a new health status fingerprint appears that does not match successfully in the library, and the effect is good after modification by a default strategy such as applying a general moving average or low-pass filtering to smooth the prediction value, then this new fingerprint and its effective correction strategy can be added to the deviation mode library as a new state mode, realizing the self-learning growth of the knowledge base.

[0077] Exemplarily, according to the result matched by the last-stage gas turbine power prediction system, such as the correction strategy matched to "positive bias compensation", dynamic compensation will be applied to the subsequently generated preliminary power prediction value. Assuming that the preliminary power value 15 minutes later at the current time is predicted to be 88.5 MW. According to the "positive bias compensation" strategy matched, the strategy stipulates that a fixed bias amount, such as 3.0 MW, should be subtracted from the preliminary prediction value, because this mode represents that the model systematically overestimates the power. Therefore, the corrected power prediction value is preliminarily calculated as 88.5 MW-3.0 MW=85.5 MW. After the dynamic compensation is implemented, the 85.5 MW compensated prediction value is immediately subjected to a reasonableness check. The check rules include checking whether the value is within the physical operating limits of the gas turbine, such as the rated maximum power of 100 MW and the minimum stable operating power of 50 MW; at the same time, checking whether its change rate exceeds the physical ramp rate limit of the device, such as a maximum ramp of 5 MW per minute. Assuming that the prediction value 85.5 MW is between 50 MW and 100 MW, and its change rate is within the physical limit, the check passes. After the check passes, the value 85.5 MW is officially output as the corrected power prediction value. If the check fails, for example, the compensated prediction value becomes 105 MW, which exceeds the maximum rated power of 100 MW, it will be determined that this compensation is unreasonable, and it will be rolled back to the uncompensated state, i.e. giving up the result of 105 MW. Then the correction strategy selection step will be triggered again, which may try the strategy corresponding to the second highest similarity score, or record this failed case for subsequent learning. The corrected power prediction value 85.5 MW that finally passes the check will be output as the final prediction result to the user and the downstream scheduling system. At the same time, the update of the bias mode library is carried out in synchronization. If the "positive bias compensation" this time significantly improves the prediction accuracy, the weight association between the "positive bias compensation" mode and the health status fingerprint type will be increased to strengthen the recognition of the effectiveness of this bias mode and correction strategy in the library. By intelligently applying the dynamic compensation strategy matched with the bias mode, the accuracy of the preliminary power prediction value can be improved, and the unusable data caused by excessive correction can be reduced, ensuring the practicality of the output result.

[0078] Optionally, the method further comprises:

[0079] obtaining a subsequent prediction bias value of the corrected power prediction output and the actual measured power value;

[0080] updating the health status fingerprint based on the subsequent prediction bias value, and performing monitoring of the performance;

[0081] when the performance indicated continues to decline, triggering a full re-calibration of the preliminary prediction model.

[0082] Specifically, the process starts with obtaining the revised power prediction value that has been dynamically compensated and outputted, and the actual measured power value corresponding to the same time point. By calculating the difference between the two, a new error sequence, i.e. the subsequent prediction bias value, is obtained. This bias value reflects the residual error of the prediction system after online revision. Next, based on this series of subsequent prediction bias values, a health status fingerprint representing the performance of the revised system is generated. This series of performance health status fingerprints is continuously generated and recorded, and this process is the performance indication monitoring. To quantify the performance indication, a comprehensive performance index can be constructed, and the higher the value of the comprehensive performance index, the better the effect of the revision. A weighted reciprocal sum can be used, and the inputs include the bias trend slope calculated based on the subsequent prediction bias values, the bias persistence positive and negative, and the fluctuation frequency. For the calculation of the comprehensive performance index , there are:

[0083]

[0084] wherein, is the bias trend slope; is the quantification of bias persistence positive and negative, such as the length or number of times of being positive or negative in succession, converted into a numerical value; is the weight coefficient, and the sum of the three is 1, which can be determined according to the regression analysis of historical data or expert experience; is a small positive number such as The smaller the numerical value of these input indicators, the more stable and unbiased the residual error is, so that the value of the comprehensive performance index is larger. The core of the monitoring process is to determine whether the performance indication is continuously decreasing. A performance warning threshold and an observation time window will be set. When the calculated comprehensive performance index is lower than the preset warning threshold for a plurality of time points in succession, or the moving average value thereof presents a significant downward trend within the observation time window, it is determined that the performance indication is continuously decreasing. This condition indicates that the current dynamic compensation strategy based on the bias pattern library has failed to effectively suppress the prediction bias, and the preliminary prediction model itself may have undergone fundamental performance degradation or has become unsuitable for the current data distribution. Once this condition is triggered, a full recalibration of the preliminary prediction model will be automatically triggered. For example, Figure 3 ​​​​As shown, the solid black line represents the overall performance index, which represents the overall performance of the corrected power prediction system over a period of time; the dashed line represents the performance warning threshold. Once the overall performance index falls below this performance warning threshold, it indicates that the performance has begun to be unsatisfactory; when it is detected that the overall performance index not only falls below the performance warning threshold, but this downward trend continues for a period of time, a full recalibration is triggered, which is the black dotted line in the figure.

[0085] For example, after outputting a corrected power prediction value, the gas turbine power prediction system continuously monitors the performance of the entire system. Suppose that at a certain point in time, the corrected power prediction output is 85.5 MW, while the actual measured power value is 86.0 MW. The difference between these two values, i.e., the subsequent prediction deviation value, is 0.5 MW. These subsequent prediction deviation values ​​are continuously collected, forming a new sequence, for example, over the past 60 minutes. Based on this new sequence of subsequent prediction deviation values, its dynamic features are extracted using the same method as generating a health status fingerprint, and a new health status fingerprint is generated. This new fingerprint no longer only evaluates the performance of the initial model, but rather assesses the overall indicative performance of the entire adaptive correction framework after dynamic compensation. For instance, a comprehensive performance index is constructed and continuously monitored. A performance warning threshold of 0.5 is set, and the observation time window is 5 consecutive sampling points. If, within a certain time period, the comprehensive performance index value is [0.45, 0.48, 0.42, 0.39, 0.41] for 5 consecutive sampling points, all below 0.5, it is determined that the indicative performance is continuously declining. This indicates that the current preliminary prediction model may be outdated, or that the correction strategies in the bias pattern library are no longer fully effective. When this condition of continuous decline is triggered, a full recalibration of the preliminary prediction model will be automatically initiated. By continuously monitoring and reanalyzing the corrected prediction bias, a high-level performance evaluation and self-diagnostic mechanism is constructed, reducing the likelihood of prolonged low prediction accuracy caused by local correction failures.

[0086] Optionally, triggering a full recalibration of the initial prediction model includes:

[0087] During the recalibration process, the historical operating data, the real-time operating conditions, and the environmental influencing factors are integrated to construct a global training sample set;

[0088] Based on the global training sample set, a preset incremental learning algorithm is used to reconstruct the structure and reset the parameters of the preliminary prediction model to generate a new generation prediction model.

[0089] The next-generation prediction model is deployed, and the health status fingerprint is reset with the deviation pattern library.

[0090] Specifically, when the recalibration process is triggered, the first step is data integration. This involves merging accumulated historical operational data with all real-time operating conditions and environmental influencing factor data continuously collected since the last calibration. The goal of this operation is to construct a global training sample set covering the entire lifecycle of the device from initial operation to its current state. This global training sample set provides the model with the most comprehensive learning material. Next, based on this global training sample set, a pre-defined incremental learning algorithm is used to upgrade the existing preliminary prediction model. Incremental learning is an advanced training strategy that efficiently incorporates new information and patterns from new data while retaining the effective knowledge learned from historical data, rather than training from scratch. During this process, the incremental learning algorithm restructures the data according to the complexity of the new data, such as adjusting the number of layers or neurons in the neural network to better match the changed data distribution. Simultaneously, all parameters of the initial prediction model are reset and re-optimized, i.e., parameter reset. Through iterative training on the global training sample set, the initial prediction model reaches a new optimal state, resulting in a new generation prediction model. During training, the model structure is reconstructed according to the complexity of the data distribution, such as dynamically adjusting the number of layers or neurons in the neural network. All learnable parameters of the model, including the weights, biases, and attention mechanism parameters of each layer, are reset. The iteration ends when the model's performance on the validation set reaches the preset convergence criterion. Finally, after the new generation prediction model is evaluated offline and its predictive performance is confirmed to be superior to the old model, deployment and reset operations are performed. First, the new generation prediction model is deployed online, replacing the original, degraded initial prediction model, making it the main force for online power prediction. Crucially, the health status fingerprint generation process and the bias pattern library related to the correction mechanism are simultaneously reset. This is because the new generation prediction model has completely different error characteristics and bias patterns from the old model, and the original bias pattern library is no longer relevant to the new model. Therefore, it is necessary to clear or archive the old deviation pattern library, so that, starting from scratch, based on the actual performance of the new generation of predictive models, we can relearn and accumulate its own health status fingerprints and deviation patterns, thereby starting a brand-new self-learning and correction cycle.

[0091] For example, when the performance of a gas turbine power prediction system continues to decline, triggering a comprehensive recalibration of the initial prediction model, a deep update process will be initiated. First, all available data is integrated. This includes all historical operating data accumulated since the equipment was put into operation, as well as all real-time operating condition data and environmental influencing factor data continuously collected since the last model calibration. For instance, all minute-level and hourly sensor data, equipment log information, and environmental meteorological data from the past three years will be collected, forming a large global training sample set. Next, based on this global training sample set, a pre-defined incremental learning algorithm, such as an online gradient descent algorithm based on deep learning, is used to restructure and reset the parameters of the initial prediction model. During this process, the differences in the distribution of old and new data are first analyzed. If a significant increase in data complexity is found, the algorithm may automatically increase the number of layers in the LSTM network or the number of neurons per layer for structural reconstruction. Simultaneously, all learnable parameters within the model, including the weights, biases, and attention mechanism parameters of each layer, are reset and comprehensively optimized on the new global training sample set. This iterative training process ensures that the model can fully learn the complex patterns contained in all historical and latest data. Once training is complete, the model, after structural reconstruction and parameter reset, becomes the next-generation prediction model with superior performance. Finally, after confirming that the prediction accuracy of the new-generation model on the offline test set is significantly better than the old model, it is deployed online, replacing the original preliminary prediction model, allowing it to immediately begin processing real-time data and generating preliminary power predictions. Crucially, historical data of the health status fingerprint associated with the new model is simultaneously cleared, and the deviation pattern library is reset. This is because after learning from all historical data and undergoing comprehensive optimization, the error characteristics and deviation patterns of the new-generation model may be completely different from the old model. Clearing the old deviation pattern library prevents old error diagnosis rules from interfering with the new model, allowing it to start from scratch, relearning and constructing its own health status fingerprint features and deviation patterns based on its actual operating performance, thus initiating a completely new self-learning and correction cycle. By integrating comprehensive historical and latest data and employing advanced incremental learning algorithms for structural reconstruction and parameter reset, the model can adapt to long-term changes in equipment operating conditions, aging, and environmental factors.

[0092] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an online update system for a self-learning and corrective power prediction model, the system comprising:

[0093] The power prediction module is used to acquire the input dataset of real-time operating conditions and environmental influencing factors related to the target power, construct a preliminary prediction model, and make predictions on the input dataset based on the preliminary prediction model to generate preliminary power prediction values.

[0094] The deviation analysis module is used to obtain and compare the actual measured power value corresponding to the preliminary power prediction value at the same time, calculate the prediction deviation value, extract the dynamic features of the prediction deviation value, and generate a health status fingerprint.

[0095] The strategy matching module is used to match the health status fingerprint with a preset deviation pattern library, determine the correction strategy, and generate a matching result.

[0096] The prediction correction module is used to trigger the corresponding correction strategy based on the matching result, and adjust the subsequently generated preliminary power prediction value to generate a corrected power prediction value.

[0097] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0098] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for online updating a power prediction model with self-learning correction, characterized in that, The method includes: Obtain the input dataset of real-time operating conditions and environmental influencing factors related to the target power, construct a preliminary prediction model, and make predictions on the input dataset based on the preliminary prediction model to generate preliminary power prediction values; Obtain and compare the actual measured power value corresponding to the preliminary power prediction value at the same time, calculate the prediction deviation value, extract the dynamic features of the prediction deviation value, and generate a health status fingerprint. The health status fingerprint is matched with a preset deviation pattern library to determine the correction strategy and generate a matching result. Based on the matching result, the corresponding correction strategy is triggered, and the subsequently generated preliminary power prediction value is adjusted to generate a corrected power prediction value.

2. The online update method for a self-learning and corrected power prediction model according to claim 1, characterized in that, The generation of preliminary power prediction values ​​includes: The system collects real-time operating conditions of the equipment, including operating load, start / stop status, equipment runtime, and operating mode switching information. It also acquires environmental influencing factors, including data on temperature, humidity, wind speed, atmospheric pressure, and solar radiation intensity, to form an input dataset. A preliminary prediction model is constructed, and dynamic feature extraction and trend fitting are performed on the input dataset based on the preliminary prediction model to generate preliminary power prediction values.

3. The online update method for a self-learning and corrected power prediction model according to claim 1, characterized in that, The construction of the preliminary prediction model includes: Based on the historical operating data of the target device and the input dataset, a prediction structure with dynamic response capability is constructed. Based on the prediction structure, multi-scale feature fusion is performed on the input dataset to capture the intrinsic relationship between short-term fluctuations and long-term trends, and to establish a mapping relationship from the real-time operating conditions and environmental influencing factors to power output. The mapping relationship is adaptively weighted and the contribution is dynamically adjusted according to the real-time strength of the influence of each element on the power in the input dataset to obtain an optimized feature expression. Through an iterative optimization process, the parameters of the prediction structure are tuned based on the optimized feature expression to obtain a preliminary prediction model.

4. The online update method for a self-learning and corrected power prediction model according to claim 1, characterized in that, The generation of the health status fingerprint includes: Based on the predicted deviation values, feature extraction is performed to obtain a sequence of historical predicted deviation values; Perform trend analysis on the historical prediction deviation value sequence to generate the slope of the deviation change trend; Statistical analysis is performed on the historical prediction deviation value sequence to generate the duration of the deviation being positive or negative and the fluctuation frequency. A health status fingerprint is generated by combining the slope of the deviation change trend, the duration of the deviation being positive or negative, and the fluctuation frequency.

5. The online update method for a self-learning and corrected power prediction model according to claim 4, characterized in that, The method further includes: Obtain the device operation status log, perform correlation analysis between the health status fingerprint and the device operation status log, and identify abnormal fingerprint patterns related to operation mode switching and load mutation. Based on the abnormal fingerprint pattern, a device health warning signal is generated, and a fine-tuning of local parameters for the preliminary prediction model is triggered.

6. The online update method for a self-learning and corrected power prediction model according to claim 4, characterized in that, The generated matching results include: Multiple state patterns are read from the deviation pattern library, wherein the state pattern is jointly characterized by the slope of the deviation change trend, the positive or negative nature of the deviation, and the fluctuation frequency; Calculate the similarity between the health status fingerprint and each of the status patterns, and generate a similarity score set; The state patterns are filtered based on the similarity score set to obtain matching results.

7. The online update method for a self-learning and corrected power prediction model according to claim 3, characterized in that, The generated corrected power prediction values ​​include: Based on the matching result, the type of the corresponding correction strategy is determined, and dynamic compensation is applied to the preliminary power prediction value, wherein the dynamic compensation includes trend compensation, bias compensation, or fluctuation suppression. After implementing the dynamic compensation, the reasonableness of the compensated prediction value is checked. If the check passes, the corrected power prediction value is output; otherwise, it reverts to the uncompensated state and the correction strategy selection step is triggered again. The corrected power prediction value is output as the final prediction result, and the deviation mode library is updated simultaneously.

8. The online update method for a self-learning and corrected power prediction model according to claim 7, characterized in that, The method further includes: Obtain the subsequent prediction deviation between the corrected power prediction output and the actual measured power value; Based on the subsequent prediction deviation value, the health status fingerprint is updated to monitor the indication performance; When the performance of the indication continues to decline, a full recalibration of the initial prediction model is triggered.

9. The online update method for a self-learning and corrected power prediction model according to claim 8, characterized in that, The triggering of a full recalibration of the initial prediction model includes: During the recalibration process, the historical operating data, the real-time operating conditions, and the environmental influencing factors are integrated to construct a global training sample set; Based on the global training sample set, a preset incremental learning algorithm is used to reconstruct the structure and reset the parameters of the preliminary prediction model to generate a new generation prediction model. The next-generation prediction model is deployed, and the health status fingerprint is reset with the deviation pattern library.

10. A self-learning and corrective online update system for a power prediction model, applied to the self-learning and corrective online update method for a power prediction model as described in any one of claims 1-9, characterized in that, The system includes: The power prediction module is used to acquire the input dataset of real-time operating conditions and environmental influencing factors related to the target power, construct a preliminary prediction model, and make predictions on the input dataset based on the preliminary prediction model to generate preliminary power prediction values. The deviation analysis module is used to obtain and compare the actual measured power value corresponding to the preliminary power prediction value at the same time, calculate the prediction deviation value, extract the dynamic features of the prediction deviation value, and generate a health status fingerprint. The strategy matching module is used to match the health status fingerprint with a preset deviation pattern library, determine the correction strategy, and generate a matching result. The prediction correction module is used to trigger the corresponding correction strategy based on the matching result, and adjust the subsequently generated preliminary power prediction value to generate a corrected power prediction value.

Images