Dynamic error introduced time series data prediction optimization method and system

Abstract

The invention belongs to the technical field of time series data prediction, and particularly relates to a dynamic error introduced time series data prediction optimization method and system. The method comprises the following steps: preprocessing to-be-processed time sequence data; inputting the preprocessed data into a pre-established and trained prediction model to obtain preliminary prediction data; comparing the preliminary prediction data with the preprocessed data to obtain a preliminary error; and based on the initial error, combining historical data of different time durations, obtaining a long-term error and a recent error according to a time decay function, and obtaining a corrected predicted value through weighted combination, thereby realizing real-time correction of the dynamic error. According to the method, the dynamic error correction method is applied to the time sequence prediction of the satellite telemetry data for the first time, the real-time correction of the prediction error is realized, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of time series data prediction, and in particular, relates to a time series data prediction optimization method and system introducing dynamic errors. Background Art

[0002] In the past few decades, time series prediction methods have been widely used in fields such as anomaly detection, event prediction, power load prediction, stock market analysis, traffic flow prediction, and health monitoring. In order to improve the prediction performance of complex time series data in practical engineering, researchers have continuously innovated and optimized various prediction methods. However, the inherent complexity of time series data stems from the nonlinearity and non-stationarity of its generation process, as well as the noise, multi-scale features, high dimensionality, missing values, and limited observation length in the data. The interaction of many factors makes time series data difficult to model and predict, which poses a severe challenge to existing analysis methods.

[0003] Traditional statistical models and machine learning algorithms have many limitations in time series forecasting, such as limited ability to extract nonlinear features, insufficient modeling of long-term dependencies, high sensitivity to data noise, and poor flexibility and applicability when processing high-dimensional data. These shortcomings lead to increased prediction errors, reduced prediction accuracy, and affected the reliability of the model in practical applications. With the rapid development of deep learning technology, it can automatically extract complex features with its multi-layer neural network structure, thereby significantly improving the accuracy and robustness in time series forecasting tasks. In particular, deep learning models such as recurrent neural networks (RNNs) and their variants (such as long short-term memory networks (LSTMs), graph neural networks (GNNs), generative adversarial networks (GANs), and Transformers have shown significant advantages in processing nonlinear relationships and long-term dependencies of time series data. However, deep learning models also have shortcomings such as high computational cost, complexity of model training and tuning, reliance on large-scale data, and poor model interpretability due to "black box" characteristics. Current time series forecasting research is developing in two main directions. One is to improve and optimize the structure of deep learning models, aiming to improve the model's fit on training data, enhance the model's expressive power, and promote the model's adaptability by optimizing data processing, model building process, and adjusting parameters.

[0004] Systematic errors often originate from factors such as the bias of the data itself, the ideality of the model assumptions, and the complexity of the external environment. Therefore, another research direction is to improve the prediction results from the perspective of error correction. This method focuses on correcting the model during the prediction process. By introducing residual networks or residual-based optimization algorithms, errors in predictions can be captured and corrected more effectively, thereby improving the accuracy of the model. However, solutions based on the perspective of error correction tend to rely on specific data sets and application scenarios, which limits their generalization ability under different conditions and leads to unstable results in new or changing environments. In addition, in actual engineering applications driven by both knowledge and data, with the dynamic changes of complex systems, the impact of historical error information on the prediction of future points is also different, which may cause the error correction-based strategy to fail to capture and update error information that changes over time, thereby limiting the comprehensiveness and accuracy of the correction effect. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a time series data prediction optimization method and system with dynamic error introduction.

[0006] In order to achieve the above object, the present invention proposes a time series data prediction optimization method introducing dynamic errors, comprising:

[0007] Step 1) preprocessing the time series data to be processed;

[0008] Step 2) inputting the preprocessed data into a pre-established and trained prediction model to obtain preliminary prediction data;

[0009] Step 3) Compare the preliminary prediction data with the preprocessed data to obtain a preliminary error;

[0010] Step 4) Based on the preliminary error, combined with historical data of different time lengths, the long-term error and the recent error are obtained according to the time decay function. Through weighted combination, the corrected prediction value is obtained to realize real-time correction of dynamic error.

[0011] Preferably, the step 1) comprises:

[0012] Data cleaning to remove outliers, fill in missing values ​​and reduce noise;

[0013] Perform feature selection on the data after data cleaning to extract key features.

[0014] Preferably, the prediction model of step 2) is PatchTST, TimesNet or iTransformer model.

[0015] Preferably, the step 3) comprises:

[0016] The actual value y of the preprocessed time series t and the initial predicted value y obtained by the prediction model t Compare and get the difference to form a sequence e t is the initial error.

[0017] Preferably, the time decay function f(h) in step 4) is exponential, fractional or Gaussian;

[0018] For exponential type, the following equation is satisfied:

[0019] f(h)=α -h

[0020] The base α is the natural base e, 2 or 3; h represents the time from the current time t to the present time;

[0021] For the fractional type, the following equation is satisfied:

[0022] f(h)=1 / 1+βh

[0023] Among them, β represents the time decay factor, which determines the decay rate of f(h);

[0024] For the Gaussian type, the following equation is satisfied:

[0025]

[0026] Where N is the sliding window size and γ is a constant value.

[0027] Preferably, the predicted value after correction in step 4) Satisfy the following formula:

[0028]

[0029] in, is the initial predicted value of the target variable y at time t, w i Represents the weighting factors of the long-term error and the recent error respectively. The sum of the two weighting factors is 1. Among them, i=1 corresponds to the p time before time t, which is the long time m, and i=2 corresponds to the p time before time t, which is the short time n, y t-h and are the true value and the preliminary predicted value of the target variable y at time th, respectively; maxerr is the maximum value of the absolute value of the difference between the true value and the preliminary predicted value at all time points within the duration of h.

[0030] Preferably, the weighting factor w i Satisfy the following formula:

[0031] w i =(1 / n)·e-λi / n ,n=2,i=1,2orw2=1-w1,w1∈[0.0,1.0]

[0032] Among them, λ is the smoothing factor.

[0033] On the other hand, the present invention also provides a time series data prediction optimization system introducing dynamic errors, comprising:

[0034] A preprocessing module is used to preprocess the time series data to be processed;

[0035] A preliminary prediction module is used to input the pre-processed data into a pre-established and trained prediction model to obtain preliminary prediction data;

[0036] A preliminary error output module is used to compare the preliminary prediction data with the preprocessed data to obtain a preliminary error;

[0037] The dynamic correction module is used to obtain the long-term error and the recent error based on the preliminary error and the historical data of different time lengths according to the time decay function, and obtain the corrected prediction value through weighted combination to realize real-time correction of dynamic error.

[0038] Compared with the prior art, the advantages of the present invention are:

[0039] 1. This invention proposes a prediction optimization method that introduces dynamic errors, adjusts the benchmark prediction results in real time, and dynamically adjusts the weights of the influence of long-term errors and recent errors on future errors based on multiple time decay functions, making the correction process more flexible and accurate. The correction results show that this method significantly improves the prediction accuracy.

[0040] 2. For the first time, the dynamic error correction method was applied to the time series prediction of satellite telemetry data, achieving real-time correction of prediction errors and verifying the robustness of the model. Combined with deep learning technology, it was verified on multiple public data sets to meet the prediction needs of complex and variable data, proving the robustness and effectiveness of the method in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Here are some examples of raw data images;

[0042] Figure 2 is the calculation process of the initial error;

[0043] Figure 3 It is a flow chart of the time series data prediction optimization method introducing dynamic error of the present invention;

[0044] Figure 4 is a visual diagram of the corrected values, preliminary predicted values ​​and true values ​​of the Electricity dataset, Exchange dataset, Weather dataset and ASOS dataset; Figure 4(a) is the benchmark prediction model using TimesNet, Figure 4(b) is the benchmark prediction model using PatchTST, and Figure 4(c) is the benchmark prediction model using iTransformer;

[0045] FIG5 is a scatter plot of the IA index values ​​of the predicted values ​​and corrected values ​​of the Electricity dataset, Exchange dataset, Weather dataset, and ASOS dataset; FIG5(a) is a baseline prediction model using TimesNet, FIG5(b) is a baseline prediction model using PatchTST, and FIG5(c) is a baseline prediction model using iTransformer;

[0046] Figure 6 is a bar chart of all index values ​​before and after DEM optimization; Figure 6 (a) is the MAE index, Figure 6 (b) is the MSE index, Figure 6 (c) is the MAPE index, Figure 6 (d) is the RMSE index, and Figure 6 (e) is the SSE index;

[0047] Figure 7 is a radar chart of all indicator values ​​of all data sets under different optimization strategies, where Figure 7(a) is the Electricity data set, Figure 7(b) is the Exchange data set, Figure 7(c) is the Weather data set, and Figure 7(d) is the ASOS data set. DETAILED DESCRIPTION

[0048] The present invention aims to provide a time series data prediction optimization method, which introduces a dynamic error correction mechanism to correct the errors in the prediction model in real time, thereby improving the accuracy and generalization performance of time series prediction, so as to meet the needs of aerospace, finance, transportation and other fields for high-precision and high-stability prediction.

[0049] include:

[0050] Step 1) preprocessing the time series data to be processed;

[0051] Step 2) inputting the preprocessed data into a pre-established and trained prediction model to obtain preliminary prediction data;

[0052] Step 3) Compare the preliminary prediction data with the preprocessed data to obtain a preliminary error;

[0053] Step 4) Based on the preliminary error, combined with historical data of different time lengths, the long-term error and the recent error are obtained according to the time decay function. Through weighted combination, the corrected prediction value is obtained to realize real-time correction of dynamic error.

[0054] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0055] Example 1

[0056] Embodiment 1 of the present invention proposes a time series data prediction optimization method introducing dynamic errors, comprising:

[0057] 1. Data preprocessing

[0058] The four data sets selected in this paper are public time series data sets from different fields: Electricity, Exchange and Weather, as well as raw telemetry data collected from a Chinese space science satellite. In order to improve the quality of the raw satellite telemetry data, data cleaning processing was carried out, including removing wild values, filling missing values, and noise reduction. Secondly, feature selection was performed on the processed telemetry data to extract key features as the input of the subsequent prediction model. Some of the raw data are as follows Figure 1 shown.

[0059] 2. Baseline prediction model

[0060] To evaluate the performance of our proposed DEM optimization method compared with current prominent time series forecasting benchmark models, including PatchTST, TimesNet, and iTransformer.

[0061] 3. Dynamic error correction mechanism

[0062] Generally speaking, the error sequence is not a simple white noise sequence, so it is believed that the error sequence still contains valuable information that can improve the prediction of the entire system. Since the error amplitude and error trend of each prediction point are closely related to the error amplitude and error trend of the previous prediction point, we can further study how to correct these errors by capturing and integrating the historical information of the error sequence to improve the prediction effect of the state of the feature variable at future moments. At the same time, considering the combined impact of long-term errors and recent errors on future errors, the dynamic adjustment of the state of the feature variable and the related issues of updating the predicted value can be realized, thereby ultimately achieving the effect of improving the prediction accuracy of the entire system.

[0063] Through in-depth analysis of long-term errors, we can better capture low-frequency information in the data (such as long-term trends and cyclical patterns) and possible systematic deviations in the model, thereby maintaining a stable error level and improving the long-term prediction ability of the model. Through in-depth analysis of recent errors, we can better capture high-frequency information in the data (such as short-term fluctuations and anomalies), thereby reducing fluctuation errors and improving the model's immediate prediction ability and accuracy.

[0064] Dynamically integrate the differential impact of long-term errors and recent errors on future errors, and consider the comprehensive estimation effects of local and global on the result state. The DEM method is proposed to weight the time dependence of the two parts of the error, dynamically adjust the system error value, and timely correct and update the target state prediction value to obtain the final prediction result. The DEM formula is calculated as follows:

[0065]

[0066] in and They are the initial forecast value and the revised forecast value of the target variable y at time t. The forecast result at time t is revised with the help of the forecast error value of the previous p periods. i ' describes the error e i Amplitude and its relative trend, y t-h and are the true value and the initial predicted value of y at time th respectively. i Represents the weighting factors of the long-term error and the recent error respectively, and the sum of the two weighting factors is 1, where i=1 corresponds to p time periods before time t, which is a long time period m, and i=2 corresponds to p time periods before time t, which is a short time period n. maxerr is the absolute value of the difference between the true value and the preliminary predicted value at all time points within the time period h, and the maximum value among them. f(h) represents the time decay function, that is, as time increases, the influence of historical information error on the predicted value at time t gradually weakens. The method in this paper includes the correction of two parts of the error sequence, namely the long-term error with an error information step length of m and the recent error with an error information step length of n. w i represents the weighting factor of each of the two parts of the error. Therefore, the proposed error correction method can dynamically utilize the historical error sequence information to achieve real-time error correction of the forecast sequence. This optimization strategy effectively solves the problems of time lag, error singleness and poor adaptability existing in traditional methods.

[0067] (1) Initial error sequence e t Calculation of: It is a sequence formed by the difference between the actual value of the time series and the preliminary predicted value obtained by the benchmark model, such as Figure 2 shown.

[0068]

[0069] (2) Dynamic error sequence e t-h Calculation of: Long-term error e t-m It refers to the difference between the predicted value and the actual value in a long period of time, reflecting the cumulative prediction error of the model over a long period of time. Similarly, the recent error e t-nIt reflects the immediate prediction error of the model in a short time span. As time goes by, the influence of long-term error information on the prediction value at time t gradually weakens. Figure 3 shown.

[0070]

[0071] (3) Time decay function f(h): exponential, Gaussian and fractional. Details are as follows:

[0072] f(h)=α -h

[0073] The α in the exponential form f(h) can be a natural base e, 2, 3, etc. In this experiment, α=e,2 is selected.

[0074] f(h)=1 / 1+βh

[0075] In the fractional type f(h), β represents the time decay factor, which determines the decay rate of f(h). In this experiment, β=0.5, 1, 2 is selected.

[0076]

[0077] In the Gaussian model f(h), N is the sliding window size, γ is a constant value, and this experiment selects γ=1, 2, 3.

[0078] (4) Weight factor w i Based on uniform exponential and linear relationships, w i It is a selective weighting of the impact of both long-term error and recent error on the forecast point. One of the following weighting methods can be used.

[0079]

[0080] Among them, λ is the smoothing factor. In this experiment, the exponential weight factor when λ=1,2 is selected.

[0081] The DEM optimization method proposed in this paper is a technology that optimizes the performance of the prediction model by fully exploring the potential valuable information contained in the error sequence. This mechanism can not only capture the instantaneous fluctuations of the error, but also identify systematic deviations and trend changes in the time dimension, thereby dynamically adjusting the model parameters. Specifically, this optimization method gradually corrects the prediction error by fusing local and global estimation results. This correction process utilizes the timely response of local recent information and the robustness of global long-term information to ensure the gradual reduction of the prediction error, and makes real-time adjustments and optimizations to future predictions, thereby significantly improving the prediction accuracy and stability. In addition, the method of introducing a dynamic error optimization prediction model can adapt to the structural changes within the model, making the model more flexible and anti-interference, especially when facing complex nonlinear and non-stationary time series data.

[0082] Example 2

[0083] Embodiment 2 of the present invention provides a time series data prediction optimization system with dynamic error introduction, which is implemented based on the method of embodiment 1 and includes:

[0084] A preprocessing module is used to preprocess the time series data to be processed;

[0085] A preliminary prediction module is used to input the pre-processed data into a pre-established and trained prediction model to obtain preliminary prediction data;

[0086] A preliminary error output module is used to compare the preliminary prediction data with the preprocessed data to obtain a preliminary error;

[0087] The dynamic correction module is used to obtain the long-term error and the recent error based on the preliminary error and the historical data of different time lengths according to the time decay function, and obtain the corrected prediction value through weighted combination to realize real-time correction of dynamic error.

[0088] Evaluation indicators:

[0089] The present invention evaluates the results from two aspects. The effectiveness of the results is tested by using MAE, MSE, MAPE, RMSE and SSE evaluation indicators. In addition, generalization ability is an important indicator for obtaining performance, and its generalization ability is reflected by the consistency index (IA).

[0090] In order to verify the accuracy, stability and generalization ability of the proposed DEM optimization method, comparative experiments were conducted on four time series data sets, three baseline models and six widely used evaluation indicators. Figure 4 shows the original values, predicted values ​​and corrected values ​​of the four data sets, namely the Electricity data set, the Exchange data set, the Weather data set and the ASOS data set, based on three benchmark prediction models. Among them, Figure 4 (a) is the benchmark prediction model using TimesNet, Figure 4 (b) is the benchmark prediction model using PatchTST, and Figure 4 (c) is the benchmark prediction model using iTransformer; Figure 5 presents the IA evaluation results of the predicted values ​​and corrected values ​​of the three benchmark models under the four data sets. Among them, Figure 5 (a) is the benchmark prediction model using TimesNet, Figure 5 (b) is the benchmark prediction model using PatchTST, and Figure 5 (c) is the benchmark prediction model using iTransformer. Figures 6 and 7 show the numerical changes of MAE, MSE, MAPE, RMSE and SSE on the four data sets, respectively, before and after the three benchmark models are optimized by the DEM method. Among them, Figure 6(a) is the MAE indicator, Figure 6(b) is the MSE indicator, Figure 6(c) is the MAPE indicator, Figure 6(d) is the RMSE indicator, and Figure 6(e) is the SSE indicator; Figure 7(a) is the Electricity dataset, Figure 7(b) is the Exchange dataset, Figure 7(c) is the Weather dataset, and Figure 7(d) is the ASOS dataset.

[0091] (1) Figure 4 intuitively shows the observed value (blue line), the preliminary predicted value (black line) and the corrected value after DEM optimization (red line). Compared with the benchmark model, the fit between the DEM optimized prediction curve and the actual observation curve is significantly improved. This visual improvement intuitively reflects the optimization effect of the DEM method, indicating that the prediction accuracy has been significantly improved and demonstrating the superior performance of this method in optimizing the benchmark model.

[0092] (2) Take a graph in Figure 5(a) as an example, which means: the scatter plot of the preliminary predicted value and the true value of the Electricity dataset under the TimesNet benchmark prediction model (blue), and the scatter plot of the corrected value and the true value after DEM correction under this benchmark prediction model (red). PreIA: the original IA value under the benchmark prediction; ModIA: the corrected IA value under this benchmark prediction model.

[0093] It can be observed from Figure 5 that compared with the IA values ​​under the previous preliminary prediction, the IA values ​​after DEM optimization are generally improved. For example, the IA values ​​of the Electricity dataset increased by 1.48%, 3.85%, and 0.54%, respectively; the IA values ​​of the Exchange dataset increased by 22.68%, 2.95%, and 2.89%, respectively. These improvements not only show the advantages of the DEM method in capturing data regularities and trends, but also prove its stability and generalization ability in different tasks.

[0094] (3) Figure 6(a)-Figure 6(e) It includes 5 3D bar graphs. Taking Figure 6(a) as an example, its meaning is: the MAE value of each dataset (Electricity, Exchange, Weather and ASOS) under 3 benchmark prediction models (TimesNet, PatchTST and iTransformer), and the MAE value of the 3 benchmark models combined with the DEM optimization method.

[0095] Tn refers to the abbreviation of TimesNet, Pa refers to the abbreviation of PatchTST; iT refers to the abbreviation of iTransformer.

[0096] Tn-Pre: preliminary value under TimesNet benchmark prediction model;

[0097] Tn-Mod: Corrected value after combining the TimesNet benchmark prediction model with the DEM method.

[0098] Same as Pa-Pre, iT-Pre, Pa-Mod, iT-Mod.

[0099] In Figures 6 and 7, the heights of the bar graphs after DEM optimization show a comprehensive decrease to varying degrees, while the closed areas in the radar graphs have become smaller and more concentrated. This shows that all indicators have improved and the overall optimization effect has been significantly improved. This result not only verifies the effectiveness of the DEM method in reducing prediction errors and improving model accuracy, but also demonstrates its robustness and wide applicability on a variety of data sets.

[0100] The present invention proposes a time series data prediction optimization method that introduces dynamic error modification (DEM). This method deeply analyzes the effective information in the error sequence, accurately identifies and captures the dynamic characteristics in the time series data, and thus optimizes the performance of the model in different time series. This method combines historical error information of different lengths: long-term errors capture low-frequency information to maintain the stability of the error, and recent errors capture high-frequency information to reduce fluctuations. By adjusting the time decay function and related parameters, real-time dynamic optimization of the prediction model is achieved. Experimental results show that when the long-term error step size is 24 and the recent error step size is 8, and the time decay function is distributed and parameterized, the overall prediction accuracy in different scenarios is improved by about 12%. This optimization strategy can effectively overcome the problems of time lag, error singleness and poor adaptability in traditional methods.

[0101] The DEM optimization method proposed in this paper significantly improves the accuracy and stability of time series prediction and effectively reduces system errors. By enhancing the generalization performance of the prediction model, it can be applied to a variety of application scenarios and data sets. This technological breakthrough provides high-precision and high-stability prediction support for aerospace, finance, transportation and other fields, and promotes the development and application of related technologies.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A time series data prediction optimization method introducing dynamic errors, comprising: Step 1) preprocessing the time series data to be processed; Step 2) inputting the preprocessed data into a pre-established and trained prediction model to obtain preliminary prediction data; Step 3) Compare the preliminary prediction data with the preprocessed data to obtain a preliminary error; Step 4) Based on the preliminary error, combined with historical data of different time lengths, the long-term error and the recent error are obtained according to the time decay function. Through weighted combination, the corrected prediction value is obtained to realize real-time correction of dynamic error.

2. The method for optimizing time series data prediction by introducing dynamic errors according to claim 1, characterized in that: The step 1) comprises: Data cleaning to remove outliers, fill in missing values ​​and reduce noise; Perform feature selection on the data after data cleaning to extract key features.

3. The time series data prediction optimization method with dynamic error introduction according to claim 1, characterized in that: The prediction model of step 2) is PatchTST, TimesNet or iTransformer model.

4. The method for optimizing time series data prediction by introducing dynamic errors according to claim 1, characterized in that: The step 3) comprises: The actual value y of the preprocessed time series t The initial prediction value obtained by the prediction model Compare and get the difference to form a sequence e t is the initial error.

5. The method for optimizing time series data prediction by introducing dynamic errors according to claim 1, characterized in that: The time decay function f(h) in step 4) is exponential, fractional or Gaussian; For exponential type, the following equation is satisfied: f(h)=a -h The base α is the natural base e, 2 or 3; h represents the time from the current time t to the present time; For the fractional type, the following equation is satisfied: f(h)=1 / 1+βh Among them, β represents the time decay factor, which determines the decay rate of f(h); For the Gaussian type, the following equation is satisfied: Where N is the sliding window size and γ is a constant value.

6. The method for optimizing time series data prediction by introducing dynamic errors according to claim 5, characterized in that: The predicted value after the correction in step 4) Satisfy the following formula: in, is the initial predicted value of the target variable y at time t, w i Represents the weighting factors of the long-term error and the recent error respectively. The sum of the two weighting factors is 1. Among them, i=1 corresponds to the p time before time t, which is the long time m, and i=2 corresponds to the p time before time t, which is the short time n, y t-h and are the true value and the preliminary predicted value of the target variable y at time th, respectively; maxerr is the maximum value of the absolute value of the difference between the true value and the preliminary predicted value at all time points within the duration of h.

7. The method for optimizing time series data prediction by introducing dynamic errors according to claim 6, characterized in that: Weighting factor w i Satisfy the following formula: w i =(1 / n)·e -λi / n ,n=2,i=1,2orw2=1-w1,w1∈[0.0,1.0] Among them, λ is the smoothing factor.

8. A time series data prediction optimization system introducing dynamic errors, characterized in that: include: A preprocessing module is used to preprocess the time series data to be processed; A preliminary prediction module is used to input the pre-processed data into a pre-established and trained prediction model to obtain preliminary prediction data; A preliminary error output module is used to compare the preliminary prediction data with the preprocessed data to obtain a preliminary error; and The dynamic correction module is used to obtain the long-term error and the recent error based on the preliminary error and the historical data of different time lengths according to the time decay function, and obtain the corrected prediction value through weighted combination to realize real-time correction of dynamic error.

Images