Air compressor energy consumption real-time prediction algorithm based on multi-sliding window mechanism

Through the integration of the multi-sliding window mechanism and multiple time series models, combined with the timing correlation between the air compressors, the problems of complexity and volatility of the air compressor energy consumption prediction are solved, and high-precision and real-time energy consumption prediction are achieved. , supports industrial energy-saving management.

CN118965150APending Publication Date: 2024-11-15SOUTH CHINA NORMAL UNIV

Patent Information

Application Number
CN202410934445.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The energy consumption prediction model of traditional air compressors has poor effect, and the energy consumption pattern is complex and changeable due to various factors, making it difficult to achieve long-term accurate prediction.

Method used

The real-time prediction algorithm for energy consumption of air compressors based on multi-sliding window mechanism is adopted, combined with ARIMA, M-ARIMA and Dlinear models, the model weight is adjusted through adaptive distribution parameter α calculation, multiple prediction results are integrated, and the timing between air compressors is considered. Relevance and data fluctuation characteristics.

Benefits of technology

It significantly reduces prediction errors, improves the accuracy and reliability of energy consumption prediction of air compressors, adapts to complex energy consumption changes, provides efficient and fine energy management solutions, and improves energy utilization efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965150A_ABST
    Figure CN118965150A_ABST
Patent Text Reader

Abstract

The invention discloses an air compressor energy consumption real-time prediction algorithm based on a multi-sliding-window mechanism. The algorithm comprises the following steps that A, data are preprocessed after being collected in real time; b, entering a multi-sliding window mechanism identification energy consumption mode through a recent sequence, entering ARIMA prediction through a historical sequence, entering M-ARIMA prediction through the historical sequence and the recent sequence, and entering Dlinear prediction through the recent sequence; c, according to ARIMA prediction, M-ARIMA prediction and Dlinear prediction, y1, y2 and y3 are output to model fusion, and model fusion is carried out after a multi-sliding-window mechanism recognition energy consumption mode is calculated through an adaptive distribution parameter alpha. By integrating multiple advanced technologies such as multiple sliding windows, M-ARIMA and Dlinear, high-precision and real-time prediction of the energy consumption of the air compressor is realized, the adaptability, the comprehensive prediction capability and the data processing strategy of the method provide powerful technical support for industrial energy saving, and the method is expected to be capable of remarkably improving the energy utilization efficiency and reducing the operation cost in practical application, and has a wide application prospect. The sustainable development of industrial production is promoted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of air compressors, and in particular to a real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism. Background Art

[0002] Air compressors (referred to as air compressors) are the main equipment for producing gas source power and are one of the indispensable equipment in industrial production activities. They are widely used in equipment manufacturing, automobiles, metallurgy, electricity, electronics, medical care, textiles and other industrial fields. In recent years, my country has continuously increased the promotion of energy-saving products. As a product used in many fields, the introduction of air compressors has great significance for my country's energy-saving cause. Traditional air compressor energy-saving methods mainly focus on the transformation of physical equipment and process flows, such as adding refrigeration equipment at the front end, transforming air handling equipment, automatically loading and unloading to control the main pipe pressure, using high-efficiency motor replacement, etc. Another method is to optimize real-time production scheduling and control strategies.

[0003] However, the above methods all rely on the improvement of physical equipment and related technologies, and have certain limitations. In addition, the air compressor system has the characteristics of time-varying, time-delayed and nonlinearity, and its energy consumption pattern is complex and changeable. The existing single model has a poor effect on the long-term prediction of air compressor energy consumption. At the same time, the energy consumption of the air compressor is affected by many factors, there are drastic fluctuations, and it shows a variety of data distribution patterns, which brings huge challenges to the prediction of long-term energy consumption. To this end, we propose a real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism. Summary of the invention

[0004] The purpose of the present invention is to provide a real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism, comprising the following steps:

[0006] A. Preprocess data after real-time data collection;

[0007] B. Enter the multi-sliding window mechanism through the recent sequence to identify the energy consumption pattern, enter the ARIMA prediction through the historical sequence, enter the M-ARIMA prediction through the historical sequence + recent sequence, and enter the Dlinear prediction through the recent sequence;

[0008] C. ARIMA prediction M-ARIMA prediction and Dlinear prediction output y1, y2 and y3 to model fusion respectively, and the multi-sliding window mechanism identifies the energy consumption mode and calculates the adaptive distribution parameter α for model fusion, and finally outputs the prediction.

[0009] Preferably, the ARIMA prediction requires the use of an ARIMA model, which is a basic model widely used in time series analysis. ARIMA is divided into two parts: AR (autoregressive model) and MA (moving average model), which respectively describe the impact of past observations and prediction errors on current observations.

[0010] in,

[0011] Y t is the time series data to be parsed, arrive Used to describe the relationship between the current value and the values ​​at p time points in the past. c is a constant term. MA: Y t =μ+∈ t +α1∈ t-1 +α2∈ t-2 +…+θ q ∈ t-q ; θ1 to θ q Used to describe the relationship between the current value and the error at the past q time points, ∈ t is the error term at time t, μ is the constant term, and the selection of p and q values ​​of the ARIMA model is done by plotting the autocorrelation function (ACF) and the partial autocorrelation function (ACF). Function, PACF), when the ACF image shows tailing and the PACF image shows truncation, the current time series is suitable for the AR model, and the lag order of the PACF truncation is the ideal value of the hyperparameter p; when the PACF image shows tailing and the ACF image shows truncation, the current time series is suitable for the MA model, and the lag order of the ACF truncation is the ideal value of the hyperparameter q; if neither the ACF nor the PACF shows obvious tailing or truncation, then the values ​​of p and q may not be 0. In this case, the Bayesian information criterion (BIC) can be used to select parameters to obtain the best fitting model under the existing data, BIC = kln(n)-21nL; where: n is the number of samples, L is the likelihood function, k is the number of model parameters, and the output BIC matrix has p as the row and q as the column.

[0012] Preferably, the Dlinear prediction needs to use the Dlinear model. In the long time series multi-step prediction task, the Dlinear model has two strategies, one is iterated multi-step (IMS), and the other is direct multi-step prediction (DMS). Multi-step iteration obtains multi-step prediction results iteratively by learning a single-step prediction model, and the prediction variance is small, but there is an effect of error accumulation. Direct multi-step prediction uses a weighted sum operation to directly use historical values ​​to predict multiple time steps, which can reduce the accumulation of prediction errors. Dlinear adopts a direct multi-step prediction mechanism.

[0013] where W∈R T×L is the learned weight matrix, L is the input sequence length, T is the output sequence length, and X i They correspond to prediction and input respectively.

[0014] Preferably, the M-ARIMA prediction requires the use of an M-ARIMA model. In actual industrial scenarios, the correlation and mutual influence between multiple air compressors are often overlooked, and this insight may provide key information on how to better perform time series prediction. The goal of this study is to use this interrelated information to improve the accuracy and performance of the prediction model. By calculating the Pearson correlation coefficient between the target air compressor and the other five air compressors, the changes in the correlation are captured and recorded in real time, thereby injecting the mutual dynamic relationship between the air compressors into the prediction model. This is accomplished by using 24 time points (representing one day) as a rolling time window. This innovative design allows the present invention to effectively mine and utilize the implicit time series association information between the air compressors. The time series correlation between the target air compressor and the other air compressors is calculated using the Pearson correlation coefficient. The specific calculation method is as follows: First, for each of the 24 data points of each day, the correlation coefficient between the target air compressor and the other five air compressors is calculated. Therefore, the correlation coefficient obtained is based on days as the time interval. The Pearson correlation coefficient:

[0015] Subsequently, the correlation coefficient of the latest day is fitted through the ARIMA model, and then the Softmax function is performed on each correlation coefficient ρi obtained by fitting, which is compressed to the range of [0, 1] and the sum of all ρi is 1. Finally, the normalized correlation coefficient is multiplied by the 24 data of other air compressors on the previous day as the M-ARIMA item prediction value of the final model. The introduction of this step is to consider the relationship between air compressors more comprehensively, adding more accurate and dynamic information to the prediction model. By carefully considering the correlation between air compressors, it is expected to improve the prediction performance of the model and make it closer to the actual scene. After obtaining the normalized correlation coefficient ρl, the prediction value of the M-ARIMA model is calculated. The formula is as follows:

[0016] y i,j refers to the predicted value of the M-ARIMA model for air compressor No. i at time j, m is the number of air compressors related to the air compressor to be predicted, Softmax(ρi) is the value of the correlation coefficient of the latest day fitted by each ARIMA model, x i,j is the energy consumption value of air compressor No. i at time j the day before, x 0,j is the energy consumption value of the air compressor to be predicted at time j the day before, k is the parameter of the correlation effect between the air compressor to be predicted and other related air compressors. The larger the k, the smaller the effect. The M-ARIMA model also considers the trend of energy consumption of the air compressor to be predicted and other related air compressors over time in the past 24 hours. Different K values ​​have a direct impact on the prediction error of the multivariate ARIMA model. For this reason, multiple K values ​​are used for comparative tests. When K=0, MAE is the largest when the energy consumption time series of the air compressor to be predicted itself in the past day is not considered, and MAE is the smallest when K=2.5. As the K value increases, the energy consumption of other related air compressors has less and less impact on the air compressor to be predicted, resulting in a gradual increase in MAE. Therefore, K is set to 2.5 in the hybrid model.

[0017] Preferably, in order to capture the time-varying distribution characteristics of time series data, a mode combining multiple sliding window sizes is designed, which can effectively characterize the time-varying characteristics of data distribution. The calculation process of the adaptive distribution parameter α is as follows:

[0018]

[0019]

[0020] The data involved in the calculation of the adaptive distribution parameter α are the first 24 points of the sequence to be predicted. S(th) is the standard deviation threshold obtained from the historical data with a sliding window size of 24. S(t) is the standard deviation of the window size of 24. S1(t) and S2(t) are the standard deviations of two groups of data divided by the 24 data with a window size of 12. S2(t) is the standard deviation of the most recent group of data. S11(t), S12(t), S21(t), and S22(t) are the standard deviations of four groups of data divided by the 24 data with a window size of 6. S22(t) is the standard deviation of the most recent group of data. Sth = m*max(St); m can be adjusted according to actual conditions.

[0021]

[0022] α1=min(S11(t) / S(th), 1).

[0023] Preferably, the model fusion uses an adaptive distribution parameter α to dynamically adjust the weight coefficient of each sub-model, and the final prediction value (Y) of the hybrid model is determined by the prediction values ​​of the three sub-models - the prediction value y1 of the ARIMA model item, the prediction value y2 of the M-ARIMA model item, the prediction value y3 of the Dlinear model item and their weight coefficients: Y = (1-β)((1-α)*y1+α*y2)+β*y3; β = min(α, (1-α)) 2 ; where the value of β is equal to min(α, (1-α)) 2 , its value will be between 0 and 0.25. The setting of α makes the predicted value Y of the hybrid model mainly depend on the predicted value of the multivariate ARIMA item when the historical time series data fluctuates greatly, and when the historical time series data tends to be stable, the predicted value Y of the hybrid model mainly depends on the predicted value of the ARIMA item. In view of the fact that the weight β of the Dlinear item is dynamically adjusted with the size of α, the influence of the predicted value of the Dlinear item on the predicted value of the hybrid model can also be properly controlled. In other words, this hybrid model can adaptively adjust the contribution of each sub-model according to the changing characteristics of the data, so as to achieve better prediction effect.

[0024] Preferably, the real-time data collection is as follows: due to the poor communication quality in complex industrial environments, and the air compressor energy consumption prediction has high requirements for data quality, the data collected by the sensor is transmitted to the server by drone relay communication, and the data preprocessing includes outlier detection and outlier processing. The outlier detection is as follows: standard deviation method: for stationary sequences, the mean and standard deviation are calculated, and the value exceeding the mean ±3σ is regarded as abnormal,

[0025] Z-score: For non-stationary sequences, the Z-score ((x-μ) / σ) of each point is calculated. Usually, data with a Z-score > 3 are considered abnormal. Outliers are processed as follows: for data with strong time series characteristics, the ARIMA model is used to predict the value of the missing point, and for values ​​with weak time series characteristics, linear interpolation is used.

[0026] Preferably, it also includes model training, multi-sliding window mechanism to identify the energy consumption distribution pattern of air compressors and hybrid model prediction, the model training is: 1) calculate the Bayesian information volume BIC to automatically select the p and q parameters of ARIMA, and then use the history of each air compressor to train the respective ARIMA model; 2) take 24 points as the window size, use the Pearson correlation coefficient to extract the time series correlation coefficient between each air compressor, and train the corresponding correlation coefficient model according to the respective correlation coefficient; 3) use the time step of the past 24 points to predict the data of the next 24 time steps, and use the historical data of each air compressor to train the corresponding Dlinear model; the multi-sliding window mechanism to identify the energy consumption distribution pattern of the air compressor is: calculate its adaptive distribution parameter α according to the real-time energy consumption data, for the subsequent model fusion; the hybrid model prediction is: adjust the weights of the three sub-models of ARIMA, Dlinear and M-ARIMA according to the size of the adaptive distribution parameter α to obtain the final energy consumption prediction data.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The invention introduces a multi-sliding window mechanism to dynamically capture the data distribution characteristics in different time periods and different granularities, so that the prediction model can be adaptively adjusted to cope with the diversification and rapid changes of the air compressor energy consumption mode. The traditional ARIMA model is combined with the multivariate extension (M-ARIMA), and the time series correlation between air compressors is effectively utilized. At the same time, the Dlinear model is introduced to strengthen the dependence on recent data, so as to construct a hybrid model framework, significantly reduce the prediction error, design advanced outlier and null value processing strategies, ensure that the prediction model is based on high-quality data, and maintain the integrity and reliability of the data through intelligent interpolation and outlier removal technology. A solution with strong versatility and high prediction efficiency is provided, which helps industrial users to implement refined management according to the prediction results, optimize the air compressor operation strategy, and realize efficient use of energy. By integrating multiple advanced technologies such as multi-sliding windows, M-ARIMA, Dlinear, etc., high-precision and real-time prediction of air compressor energy consumption is realized. Its adaptability, comprehensive prediction ability and data processing strategy provide strong technical support for industrial energy conservation. It is expected that in practical applications, it can significantly improve energy utilization efficiency, reduce operating costs, and promote the sustainable development of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1It is a flow chart of the present invention;

[0030] Figure 2 is the ACF diagram of the present invention;

[0031] Figure 3 It is the PACF diagram of the present invention;

[0032] Figure 4 It is the Dlinear graph of the present invention;

[0033] Figure 5 It is a time series correlation coefficient diagram of the present invention;

[0034] Figure 6 This is a curve diagram showing the change of MAE with K value in the present invention;

[0035] Figure 7 This is a flow chart of the adaptive distribution parameter α of the present invention;

[0036] Figure 8 Comparison of the prediction results of the present invention Figure 1 ;

[0037] Fig. 9 Comparison of the prediction results of the present invention Figure 2 ;

[0038] Fig.10 Comparison of the prediction results of the present invention Figure 3 ;

[0039] Fig.11 Comparison of the prediction results of the present invention Figure 4 ;

[0040] Fig.12 It is a comparison diagram of the hybrid model and each submodule of the present invention;

[0041] Fig.13 This is a comparison chart of other models of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] See also Figure 1-Figure 13 , a real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism, comprising the following steps:

[0044] A. Preprocess data after real-time data collection;

[0045] B. Enter the multi-sliding window mechanism through the recent sequence to identify the energy consumption pattern, enter the ARIMA prediction through the historical sequence, enter the M-ARIMA prediction through the historical sequence + recent sequence, and enter the Dlinear prediction through the recent sequence;

[0046] C. ARIMA prediction M-ARIMA prediction and Dlinear prediction output y1, y2 and y3 to model fusion respectively, and the multi-sliding window mechanism identifies the energy consumption mode and calculates the adaptive distribution parameter α for model fusion, and finally outputs the prediction.

[0047] ARIMA forecasting requires the use of the ARIMA model, which is a basic model widely used in time series analysis. ARIMA is divided into two parts: AR (autoregressive model) and MA (moving average model), which respectively describe the impact of past observations and forecast errors on current observations.

[0048] in,

[0049] Y r is the time series data to be parsed, arrive Used to describe the relationship between the current value and the values ​​at p time points in the past. c is a constant term. MA: Y t =μ+∈ t +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q ; θ1 to θ q Used to describe the relationship between the current value and the error at the past q time points, ∈ tis the error term at time t, μ is the constant term, and the selection of p and q values ​​of the ARIMA model is done by plotting the autocorrelation function (ACF) and the partial autocorrelation function (ACF). Function, PACF), when the ACF image shows tailing and the PACF image shows truncation, the current time series is suitable for the AR model, and the lag order of the PACF truncation is the ideal value of the hyperparameter p; when the PACF image shows tailing and the ACF image shows truncation, the current time series is suitable for the MA model, and the lag order of the ACF truncation is the ideal value of the hyperparameter q; if neither the ACF nor the PACF shows obvious tailing or truncation, then the values ​​of p and q may not be 0. In this case, the Bayesian information criterion (BIC) can be used to select parameters to obtain the best fitting model under the existing data, BIC = kln(n)-21nL; where: n is the number of samples, L is the likelihood function, k is the number of model parameters, and the output BIC matrix has p as the row and q as the column.

[0050] Dlinear prediction requires the use of the Dlinear model. In the long-term multi-step prediction task, the Dlinear model has two strategies: iterated multi-step (IMS) and direct multi-step prediction (DMS). Multi-step iteration iteratively obtains multi-step prediction results by learning a single-step prediction model. The prediction variance is small, but there is an impact of error accumulation. Direct multi-step prediction uses a weighted sum operation to directly use historical values ​​to predict multiple time steps, which can reduce the accumulation of prediction errors. Dlinear uses a direct multi-step prediction mechanism. where W∈R T×L is the learned weight matrix, L is the input sequence length, T is the output sequence length, and X i They correspond to prediction and input respectively.

[0051] M-ARIMA prediction requires the use of the M-ARIMA model. In actual industrial scenarios, the correlation and mutual influence between multiple air compressors are often overlooked, and this insight may provide key information on how to better perform time series forecasting. The goal of this study is to use this interrelated information to improve the accuracy and performance of the forecasting model. By calculating the Pearson correlation coefficient between the target air compressor and the other five air compressors, the changes in the correlation are captured and recorded in real time, thereby injecting the mutual dynamic relationship between the air compressors into the forecasting model. This is done by using 24 time points (representing one day) as a rolling time window. This innovative design allows the present invention to effectively mine and utilize the implicit time series association information between air compressors. The Pearson correlation coefficient is used to calculate the time series correlation between the target air compressor and the other air compressors. The specific calculation method is as follows: First, for each of the 24 data points of each day, the correlation coefficient between the target air compressor and the other five air compressors is calculated. Therefore, the correlation coefficient obtained is based on days as the time interval. The Pearson correlation coefficient: Subsequently, the correlation coefficient of the latest day is fitted through the ARIMA model, and then the Softmax function is performed on each correlation coefficient ρi obtained by fitting, which is compressed to the range of [0, 1] and the sum of all ρi is 1. Finally, the normalized correlation coefficient is multiplied by the 24 data of other air compressors on the previous day as the M-ARIMA item prediction value of the final model. The introduction of this step is to consider the relationship between air compressors more comprehensively, adding more accurate and dynamic information to the prediction model. By carefully considering the correlation between air compressors, it is expected to improve the prediction performance of the model and make it closer to the actual scene. After obtaining the normalized correlation coefficient ρl, the prediction value of the M-ARIMA model is calculated. The formula is as follows: y i,j refers to the predicted value of the M-ARIMA model for air compressor No. i at time j, m is the number of air compressors related to the air compressor to be predicted, Softmax(ρi) is the value of the correlation coefficient of the latest day fitted by each ARIMA model, x i,j is the energy consumption value of air compressor No. i at time j the day before, x 0,jis the energy consumption value of the air compressor to be predicted at time j the day before, k is the parameter of the correlation effect between the air compressor to be predicted and other related air compressors. The larger the k, the smaller the effect. The M-ARIMA model also considers the trend of energy consumption of the air compressor to be predicted and other related air compressors over time in the past 24 hours. Different K values ​​have a direct impact on the prediction error of the multivariate ARIMA model. For this reason, multiple K values ​​are used for comparative tests. When K=0, MAE is the largest when the energy consumption time series of the air compressor to be predicted itself in the past day is not considered, and MAE is the smallest when K=2.5. As the K value increases, the energy consumption of other related air compressors has less and less impact on the air compressor to be predicted, resulting in a gradual increase in MAE. Therefore, K is set to 2.5 in the hybrid model.

[0052] In order to capture the time-varying distribution characteristics of time series data, a mode combining multiple sliding window sizes is designed, which can effectively characterize the time-varying characteristics of data distribution. The calculation process of the adaptive distribution parameter α is as follows:

[0053]

[0054] The data involved in the calculation of the adaptive distribution parameter α are the first 24 points of the sequence to be predicted. S(th) is the standard deviation threshold obtained from the historical data with a sliding window size of 24. S(t) is the standard deviation of the window size of 24. S1(t) and S2(t) are the standard deviations of two groups of data divided by the 24 data with a window size of 12. S2(t) is the standard deviation of the most recent group of data. S11(t), S12(t), S21(t), and S22(t) are the standard deviations of four groups of data divided by the 24 data with a window size of 6. S22(t) is the standard deviation of the most recent group of data. Sth = m*max(St); m can be adjusted according to actual conditions.

[0055]

[0056] α1=min(S11(t) / S(th), 1).

[0057] Model fusion uses an adaptive distribution parameter α to dynamically adjust the weight coefficient of each sub-model. The final prediction value (Y) of the hybrid model is determined by the prediction values ​​of the three sub-models - the prediction value y1 of the ARIMA model item, the prediction value y2 of the M-ARIMA model item, and the prediction value y3 of the Dlinear model item and their weight coefficients: Y = (1-β)((1-α)*y1+α*y2)+β*y3; β = min(α, (1-α)) 2 ; where the value of β is equal to min(α, (1-α)) 2, its value will be between 0 and 0.25. The setting of α makes the predicted value Y of the hybrid model mainly depend on the predicted value of the multivariate ARIMA item when the historical time series data fluctuates greatly, and when the historical time series data tends to be stable, the predicted value Y of the hybrid model mainly depends on the predicted value of the ARIMA item. In view of the fact that the weight β of the Dlinear item is dynamically adjusted with the size of α, the influence of the predicted value of the Dlinear item on the predicted value of the hybrid model can also be properly controlled. In other words, this hybrid model can adaptively adjust the contribution of each sub-model according to the changing characteristics of the data, so as to achieve better prediction effect.

[0058] Real-time data collection is as follows: Due to the poor communication quality in complex industrial environments, and the air compressor energy consumption prediction has high requirements for data quality, the data collected by the sensor is transmitted to the server by drone relay communication, and the data preprocessing includes outlier detection and outlier processing. Outlier detection is as follows: Standard deviation method: For stationary sequences, the mean and standard deviation are calculated, and values ​​exceeding the mean ±3σ are considered abnormal. Z-score: For non-stationary sequences, the Z-score ((x-μ) / σ) of each point is calculated. Usually, data with Z-score>3 are considered abnormal. Outlier processing is as follows: For data with strong time series characteristics, the ARIMA model is used to predict the values ​​of missing points, and linear interpolation is used for values ​​with weak time series characteristics.

[0059] It also includes model training, multi-sliding window mechanism to identify the energy consumption distribution pattern of air compressors and hybrid model prediction. The model training is as follows: 1) Calculate the Bayesian information volume BIC to automatically select the p and q parameters of ARIMA, and then use the history of each air compressor to train their own ARIMA model; 2) Take 24 points as the window size, use the Pearson correlation coefficient to extract the time series correlation coefficient between each air compressor, and train the corresponding correlation coefficient model according to their respective correlation coefficients; 3) Use the time step of the past 24 points to predict the data of the next 24 time steps, and use the historical data of each air compressor to train the corresponding Dlinear model; The multi-sliding window mechanism identifies the energy consumption distribution pattern of the air compressor: calculate its adaptive distribution parameter α based on the real-time energy consumption data for subsequent model fusion; The hybrid model prediction is: adjust the weights of the three sub-models of ARIMA, Dlinear, and M-ARIMA according to the size of the adaptive distribution parameter α to obtain the final energy consumption prediction data.

[0060] Multi-sliding window mechanism: Feature description: This algorithm innovatively designs a multi-sliding window mechanism, which can flexibly adapt to the dynamic distribution characteristics of air compressor energy consumption data. By setting sliding windows of different sizes (such as 24 hours, 12 hours, 6 hours, etc.), the system can carefully analyze the fluctuations and trends of data on different time scales, and effectively capture the short-term and long-term changes in energy consumption patterns, thereby providing more accurate input information for the prediction model; Advantage analysis: Dynamic adaptability: Automatically identify and adapt to the time-varying characteristics of data distribution, which improves the prediction accuracy of the model under different working conditions; Comprehensive feature capture: Different window sizes cover a full range of features from detailed fluctuations to macro trends, enhancing the model's ability to learn complex patterns.

[0061] M-ARIMA multivariate prediction model: Feature description: The M-ARIMA model extends the classic ARIMA model to a multivariate environment. It not only considers the energy consumption time series of a single air compressor, but also incorporates the correlation analysis with other air compressors. By calculating the Pearson correlation coefficient in real time, it captures the interaction and influence between machines, making the prediction more in line with the synergy in the actual working scenario; Advantage analysis: Enhanced correlation analysis: The prediction weights are dynamically adjusted by using the correlation between air compressors, which improves the overall performance of the prediction model, especially in periods of large data fluctuations; Comprehensive prediction capabilities: Integrating multi-source information, it improves the prediction sensitivity to abnormal fluctuations in energy consumption and reduces the prediction bias of a single model due to ignoring interactions.

[0062] Dlinear direct multi-step prediction: Feature description: The Dlinear model optimizes the stability and accuracy of long-term predictions through a direct multi-step prediction strategy. Its unique weighted sum operation reduces the error accumulation problem in the prediction process, and is particularly suitable for processing air compressor energy consumption data with long-term dependence; Advantage analysis: Reduce prediction error accumulation: Directly predict multi-step future values ​​to avoid error propagation problems that may occur in iterative predictions; Efficient use of historical information: Strengthens the role of recent data in predictions and improves the response speed to recent trend changes.

[0063] Adaptive weight allocation: Feature description: The hybrid model dynamically adjusts the weights of the three sub-models ARIMA, M-ARIMA, and Dlinear through the adaptive distribution parameter α, ensuring that the model can automatically adjust the strategy when the data characteristics change, balancing the impact of historical trends, immediate correlations, and recent dynamics; Advantage analysis: Flexibility and robustness: The model can adaptively adjust according to real-time changes in data distribution characteristics to maintain high prediction accuracy and stability; Optimize resource allocation: Intelligently allocate model resources to avoid the decline in prediction efficiency that may be caused by fixed weight allocation when data characteristics change.

[0064] Strengthen data processing and model training: Feature description: including the use of drone relay communication to ensure data quality, the implementation of strict data preprocessing processes (such as outlier detection and processing) and automated model parameter selection (such as BIC criterion to select ARIMA parameters), ensuring that model training is based on high-quality and standardized data; Advantage analysis: Improve prediction reliability: Ensure the accuracy and completeness of data from the source, providing a solid foundation for model training; Simplify model maintenance: Automated parameter selection reduces manual intervention, improves the efficiency of model training and the convenience of model deployment.

[0065] Experimental Results

[0066] Experiment results

[0067] Comparison of prediction results under four data distribution modes

[0068] The data in this study showed four main distribution patterns. Therefore, we compared the performance of ARIMA model, Dlinear model, M-ARIMA model and hybrid model under these four data distributions. Figure 8 As shown in the figure, when the historical data fluctuates significantly, the performance of the ARIMA model, the Dlinear model, and the M-ARIMA model are all poor, while the performance of the hybrid model is excellent. The prediction results of the ARIMA model are significantly higher than the actual values, which shows that when faced with drastic fluctuations in historical data, the prediction ability of the ARIMA model will be weakened. Although the predictions of the Dlinear model also deviate significantly from the actual values, it can effectively learn some fluctuation trends by enhancing the influence of historical values. In contrast, the predictions of the M-ARIMA model are close to the actual values. It uses the time series correlation of other variables to show excellent results when processing drastically fluctuating data. There is no doubt that the hybrid model combines the advantages of the Dlinear and M-ARIMA models, and thus has the best prediction performance.

[0069] like Fig. 9 As shown in the figure, when the fluctuation of data is large in the early stage and small in the later stage and gradually tends to be stable, the M-ARIMA model performs the worst. Although it roughly predicts the trend of the actual value, the difference between the predicted values ​​of the Dlinear model and the ARIMA model and the true value is small, but no obvious trend is predicted, and the prediction results of the hybrid model are almost completely consistent with the actual value.

[0070] like Fig.10 As shown in the figure, when faced with a situation where multiple distributions of data appear alternately, the performance of the ARIMA model and the Dlinear model lags behind, the performance of the M-ARIMA model is average, and the mixed model performs best.

[0071] like Fig.11 As shown in the figure, when the data distribution is relatively stable, the performance of the M-ARIMA model and the Dlinear model is relatively weak, while the ARIMA model and the hybrid model both show good performance.

[0072] Comparison of MAE and RMSE of each model under four data distribution modes: According to the results in the table below, we can see that the hybrid model performs well in the MAE and RMSE evaluation criteria under all four data distribution modes. The ARIMA model performs particularly well when the data distribution is stable. However, its predictive ability is significantly reduced when the data shows fluctuations. The Dlinear model can show ideal results when the data fluctuations are not large, while the M-ARIMA model shows excellent predictive ability when dealing with data fluctuations.

[0073] Prediction Result in different data type

[0074]

[0075]

[0076]

[0077] Ablation experiment: In order to accurately evaluate the contribution of various models to the hybrid model, this paper conducted four sets of comparative experiments. The results are as follows: Fig.12 As shown in the figure, the MAE and RMSE of the single ARIMA model are 176.2 and 229.0; when the Dlinear model is introduced, the predicted MAE and RSME are increased by 7.2% and 7.1% respectively; when the M-ARIMA model is introduced, the predicted MAE and RSME are increased by 41.9% and 49.3% respectively. Finally, when the Dlinear model and the M-ARIMA model are introduced at the same time, the MAE and RSME of the hybrid model are increased by 55.6% and 61.2% respectively. These results show that although the Dlinear model has a certain improvement on the hybrid model, its influence is not great. This is mainly because the data of air compressors generally have large volatility, and the Dlinear model has a limited ability to handle such volatile data. On the other hand, the M-ARIMA model has a significantly greater improvement on the hybrid model, thanks to its ability to extract time series correlations from other air compressors, enabling it to more effectively handle data with large volatility.

[0078] Comparison with other models: In order to compare the performance of other mainstream models on the current dataset, we conducted additional control experiments, such as Fig.13As shown in the figure, the performance of VAR and LSTNet models is similar to that of ARIMA model, while the model based on Transformer architecture performs relatively poorly on this dataset. However, when comparing the hybrid model with these models, it can be seen that the former is significantly better than the latter. The reason why the hybrid model performs so well is that it integrates the advantages of multiple models and can dynamically adjust the weights of sub-models according to different data distribution patterns, which makes it far superior to other single time series prediction models in terms of MAE and RSME evaluation criteria. Specifically, in terms of MAE indicator, the hybrid model is 88.4, 100.7, 103.5, and 120.7 lower than VAR, LSTNet, Informer, and FEDformer, while in terms of RSME indicator, the hybrid model is 106.9, 136.6, 171.2, and 180.1 lower than VAR, LSTNet, Informer, and FEDformer.

[0079] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism, characterized by: The following steps are involved: A. Preprocess data after real-time data collection; B. Enter the multi-sliding window mechanism through the recent sequence to identify the energy consumption pattern, enter the ARIMA prediction through the historical sequence, enter the M-ARIMA prediction through the historical sequence + recent sequence, and enter the Dlinear prediction through the recent sequence; C. ARIMA prediction M-ARIMA prediction and Dlinear prediction output y1, y2 and y3 to model fusion respectively, and the multi-sliding window mechanism identifies the energy consumption mode and calculates the adaptive distribution parameter α for model fusion, and finally outputs the prediction.

2. According to claim 1, a real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism is characterized in that: The ARIMA forecast requires the use of the ARIMA model, which is a basic model widely used in time series analysis. ARIMA is divided into two parts, AR and MA, which respectively describe the impact of past observations and prediction errors on current observations. Among them, Y t is the time series data to be parsed, arrive It is used to describe the relationship between the current value and the value at p time points in the past. c is a constant term. MA:Y t =μ+∈ t +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q ; θ1 to θ q Used to describe the relationship between the current value and the error at the past q time points, ∈ t is the error term at time point t, and μ is the constant term.

3. The real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism according to claim 1 is characterized in that: The Dlinear prediction requires the use of the Dlinear model. In the long-term multi-step prediction task, the Dlinear model has two strategies: multi-step iteration and direct multi-step prediction. Multi-step iteration iteratively obtains multi-step prediction results by learning a single-step prediction model. The prediction variance is small, but there is an impact of error accumulation. The direct multi-step prediction uses a weighted sum operation to directly use historical values ​​to predict multiple time steps, which can reduce the accumulation of prediction errors. Dlinear uses a direct multi-step prediction mechanism. where W∈R T×L is the learned weight matrix, L is the input sequence length, T is the output sequence length, and X i They correspond to prediction and input respectively.

4. The real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism according to claim 1 is characterized in that: The M-ARIMA prediction requires the use of the M-ARIMA model. By calculating the Pearson correlation coefficient between the target air compressor and the other five air compressors, the changes in correlation are captured and recorded in real time, thereby injecting the dynamic relationship between the air compressors into the prediction model. This is done by using 24 time points as rolling time windows. The Pearson correlation coefficient is used to calculate the time series correlation between the target air compressor and the other air compressors. The specific calculation method is as follows: First, for each of the 24 data points of each day, the correlation coefficient between the target air compressor and the other five air compressors is calculated. Therefore, the correlation coefficient is obtained with days as the time interval. The Pearson correlation coefficient is: Subsequently, the correlation coefficient of the latest day is fitted through the ARIMA model, and then the Softmax function is performed on each correlation coefficient ρi obtained by fitting, which is compressed to the range of [0, 1] and the sum of all ρi is 1. Finally, the normalized correlation coefficient is multiplied by the 24 data of other air compressors on the previous day as the M-ARIMA item prediction value of the final model. After obtaining the normalized correlation coefficient ρi, the prediction value of the M-ARIMA model is calculated. The formula is as follows: y i,j refers to the predicted value of the M-ARIMA model for air compressor No. i at time j, m is the number of air compressors related to the air compressor to be predicted, Softmax(ρi) is the value of the correlation coefficient of the latest day fitted by each ARIMA model, x i,j is the energy consumption value of air compressor No. i at time j the day before, x 0,j is the energy consumption value of the air compressor to be predicted at time j the day before, k is the parameter of the correlation between the air compressor to be predicted and other related air compressors. The larger the k, the smaller the impact. The M-ARIMA model also considers the trend of energy consumption of the air compressor to be predicted and other related air compressors over time in the past 24 hours.

5. The real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism according to claim 1 is characterized in that: The calculation process of the adaptive distribution parameter α is as follows: The data involved in the calculation of the adaptive distribution parameter α are the first 24 points of the sequence to be predicted. S(th) is the standard deviation threshold obtained from the historical data with a sliding window size of 24. S(t) is the standard deviation of the window size of 24. S1(t) and S2(t) are the standard deviations of two groups of data divided by the 24 data with a window size of 12. S2(t) is the standard deviation of the most recent group of data. S11(t), S12(t), S21(t), and S22(t) are the standard deviations of four groups of data divided by the 24 data with a window size of 6. S22(t) is the standard deviation of the most recent group of data. Sth = m*max(St); m can be adjusted according to actual conditions. α1=min(S11(t) / S(th), 1).

6. The real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism according to claim 1 is characterized in that: The model fusion uses an adaptive distribution parameter α to dynamically adjust the weight coefficient of each sub-model. The final prediction value (Y) of the hybrid model is determined by the prediction values ​​of the three sub-models - the prediction value y1 of the ARIMA model item, the prediction value y2 of the M-ARIMA model item, and the prediction value y3 of the Dlinear model item and their weight coefficients: Y = (1-β)((1-α)*y1+α*y2)+β*y3; β = min(α, (1-α)) 2 ; where the value of β is equal to min(α, (1-α)) 2 , its value will be between 0 and 0.

25. The setting of α makes the predicted value Y of the hybrid model mainly depend on the predicted value of the multivariate ARIMA item when the historical time series data fluctuates greatly, and the predicted value Y of the hybrid model mainly depends on the predicted value of the ARIMA item when the historical time series data tends to be stable.

7. The real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism according to claim 1 is characterized in that: The real-time data collection is as follows: the data collected by the sensor is transmitted to the server by means of drone relay communication, and the data preprocessing includes outlier detection and outlier processing. The outlier detection is as follows: standard deviation method: for stationary series, the mean and standard deviation are calculated, and the values ​​exceeding the mean ±3σ are regarded as abnormal. Z-score: for non-stationary series, the Z-score ((x-μ) / σ) of each point is calculated. Usually, data with |Z-score|>3 are regarded as abnormal. Outlier processing is as follows: for data with strong time series characteristics, the ARIMA model is used to predict the values ​​of missing points, and linear interpolation is used for values ​​with weak time series characteristics.

8. The real-time prediction algorithm for air compressor energy consumption based on a multi-sliding window mechanism according to claim 1 is characterized in that: It also includes model training, multi-sliding window mechanism to identify the energy consumption distribution pattern of air compressors and hybrid model prediction. The model training is as follows: 1) Calculate the Bayesian information volume BIC to automatically select the p and q parameters of ARIMA, and then use the history of each air compressor to train their own ARIMA model; 2) Take 24 points as the window size, use the Pearson correlation coefficient to extract the time series correlation coefficient between each air compressor, and train the corresponding correlation coefficient model according to their respective correlation coefficients; 3) Use the time step of the past 24 points to predict the data of the next 24 time steps, and use the historical data of each air compressor to train the corresponding Dlinear model; The multi-sliding window mechanism identifies the energy consumption distribution pattern of the air compressor: calculate its adaptive distribution parameter α based on the real-time energy consumption data for subsequent model fusion; The hybrid model prediction is: adjust the weights of the three sub-models of ARIMA, Dlinear, and M-ARIMA according to the size of the adaptive distribution parameter α to obtain the final energy consumption prediction data.

Citation Information

Patent Citations

  • Method for predicting short-time passenger flow of bus

    CN104517159A

  • Multivariable time series data prediction model based on double-window mechanism

    CN113392137A

  • Industrial production index prediction method and processor

    CN115271192A

  • Short-term load prediction method based on SARIMA-random forest combination model

    CN115800245A

  • Neural network model, time sequence prediction method and device, and storage medium

    CN117454942A

Cited By

  • Intelligent monitoring method, device and equipment for machine room power and storage medium

    CN119493708A

  • Smart factory energy efficiency optimization method and system, electronic equipment and storage medium

    CN119990413A

  • A smart factory energy efficiency optimization method, system, electronic device and storage medium

    CN119990413B

  • A dynamic error real-time compensation control method and system of a high-speed motion system

    CN122794829A