Enclosed busbar wind speed prediction control method based on space-time intelligent adaptation prediction engine
Through deep learning and feature extraction technology based on space-time intelligent adaptation prediction engine, the existing fan wind speed prediction methods are solved, and high accuracy wind speed prediction and dynamic adjustment are achieved, and energy efficiency and Equipment life, while improving the working environment and system robustness.
Patent Information
- Application Number
- CN202411570147.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-06
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing fan wind speed prediction methods have problems such as overfitting, high computational complexity, sensitivity to missing data and noise, and inability to process nonlinear time series data, resulting in low prediction accuracy, especially in complex systems, which is difficult to achieve accurate wind speed prediction. .
Using a method based on space-time intelligent adaptation prediction engine, through deep learning and feature extraction technology, a basic neural network and space analysis engine are established, combined with time-series analysis engine, integrated deep space analysis and time-series analysis to form a space-time intelligent adaptation prediction algorithm for wind speed Predict and control.
It realizes high-precision wind speed prediction, dynamic adaptive adjustment, improves energy efficiency, extends equipment life, improves working environment, and enhances system robustness, supporting decision-making.
Smart Images

Figure CN119089410A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wind speed prediction, and specifically relates to a closed bus wind speed prediction and control method based on a spatiotemporal intelligent adaptive prediction engine. Background Art
[0002] A closed busbar is a cable layout used in a power system, mainly used to connect power equipment such as generators, transformers, and fans. Closed busbar fans are usually used to control the air flow in a closed busbar system to ensure equipment heat dissipation and stable system operation. A fan is a device that converts mechanical energy into air kinetic energy. It is widely used in industry, construction, and power systems for the transportation, ventilation, cooling, and heating of air or gas. The stable operation of the fan is critical to the safety and reliability of the equipment. Changes in wind speed may affect the working state and performance of the fan, so it is necessary to accurately predict the wind speed in order to take preventive measures in advance. Existing methods for predicting wind speed for fans include artificial neural networks, support vector machines, gray prediction models, Kalman filters, and other methods, but these methods are all insufficient and prone to various problems in practical applications.
[0003] Artificial neural networks require a large amount of historical data for training, otherwise overfitting may occur. Training complex neural networks may take a lot of time, especially for large-scale networks, which requires a long time and high computing resources. In addition, the decision-making process within the neural network lacks transparency. Support vector machine SVM needs to select appropriate kernel functions and parameters, otherwise it may affect the prediction accuracy; in addition, SVM has high computational complexity and requires a long time for training and prediction; SVM is sensitive to missing data and noise, which may affect the prediction accuracy; selecting appropriate kernel functions and regularization parameters usually requires cross-validation, which is a time-consuming process. SVM has high computational cost when processing large data sets because it needs to solve a quadratic programming problem. Grey prediction models need to preprocess historical data, otherwise it may affect the prediction accuracy; grey prediction models cannot handle nonlinear time series data, so when dealing with complex wind speed changes, the prediction accuracy may be affected; in addition, grey prediction models perform better in short-term predictions, and the accuracy of long-term predictions will decrease. Usually, at least 4 data points are required, and the data must show a certain regularity. The Kalman filter needs to estimate the state of wind speed, otherwise it may affect the prediction accuracy; the computational complexity of the Kalman filter is high and it takes a long time to make predictions; the Kalman filter is sensitive to the estimation of the initial state, and a poor initial estimation may cause the filter to diverge. For complex systems, the computational complexity of the Kalman filter may be very large. Summary of the invention
[0004] In view of the above deficiencies in the prior art, the purpose of the present invention is to provide a closed bus wind speed prediction and control method based on a spatiotemporal intelligent adaptive prediction engine, which can predict the wind speed of wind turbines inside and outside the closed bus, and can adjust the wind speed of the wind turbines in time according to the control strategy.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is a closed bus wind speed prediction and control method based on a spatiotemporal intelligent adaptive prediction engine, comprising the following steps: S1. For the fans at the closed busbar where wind speed prediction is required, wind speed thresholds are set, including high wind speed thresholds and low wind speed thresholds, and a control system is configured. The control system is used to adjust the fan wind speed in combination with the wind speed prediction and the closed busbar temperature obtained by real-time monitoring to ensure that the closed busbar temperature is within a safe range; S2. For wind turbines that need wind speed prediction, collect historical data of wind speed of each wind turbine, process missing values and abnormal values in the historical data, normalize and smooth the processed historical data, complete preprocessing of the historical data, and divide the preprocessed historical data into a training set and a test set; S3. Extract features from the preprocessed historical data based on deep learning and analyze the time series features, which include trends, seasonality, and periodicity. S4. Establish a basic neural network architecture for processing time series features. The basic neural network architecture includes three parts: input layer, hidden layer, and output layer. By connecting the output of the hidden layer to the input of the next time point, a loop structure is formed to capture the time dependency in the time series features. S5. Establish a spatial analysis engine, including a forget gate, an input gate, a memory unit, and an output gate, to optimize the output of the basic neural network architecture; S6. Establish a deep spatial analysis engine. The deep spatial analysis engine consists of multiple layers of spatial analysis engines and can capture more complex data features. S7, constructing a prediction method based on a deep spatial analysis engine, the steps including data preparation, building a deep spatial analysis engine model, model training, wind speed prediction and loss calculation; S8. Establish a time series analysis engine, including autoregression, difference and moving average, for predicting time series data with autoregression and moving average characteristics; S9, constructing a prediction method based on a time series analysis engine, the steps including defining model parameters, defining a fitness function, calculating fitness, selecting the best, and setting a termination condition; S10, integrating the deep spatial analysis engine with the time series analysis engine to obtain a spatiotemporal intelligent adaptive prediction engine algorithm; S11. Apply the spatiotemporal intelligent adaptive prediction engine algorithm to perform time series prediction of wind speed of the wind turbine to obtain the final predicted value of the wind speed; S12. Based on the final predicted value of the wind speed and the closed bus temperature obtained through real-time monitoring, a control strategy for the corresponding fan is formulated, and the fan speed is adjusted through the control system according to the control strategy.
[0006] As a preferred solution of the present invention, in S2, a wind speed sensor is used to collect historical data of wind speed of each wind turbine, and missing values are supplemented by interpolation and mean filling; outliers are identified by using standard scores, and replaced or deleted; When smoothing data, single exponential smoothing is used for data without trend; double exponential smoothing is used for data with trend but no seasonality; and triple exponential smoothing is used for data with trend and seasonality.
[0007] As a preferred embodiment of the present invention, in S3, for trend analysis, based on the extracted features, a time series diagram of wind speed data is drawn, and a trend line is fitted to the wind speed time series data using the least squares curve fitting technique to describe the trend in the form of a mathematical model; For seasonal analysis, the extracted features were grouped by season, and the average wind speed for each season was plotted to see if there was an obvious seasonal pattern. The wind speed differences between different seasons were compared to analyze the impact of seasonality on wind speed. For periodicity analysis, the autocorrelation function is calculated and the ACF plot is used to analyze the correlation of the features extracted from wind speed at different time lags to find periodic patterns; check whether the period remains stable over time and whether there is a trend of gradual increase or decrease; In addition, combined with the closed busbar temperature obtained by real-time monitoring, a joint trend analysis of wind speed and temperature characteristics was performed to observe whether temperature changes had a significant impact on wind speed.
[0008] As a preferred solution of the present invention, in S5, the forget gate determines how much past information is retained, and calculates the amount of information of the previous state through a sigmoid function: ; In the formula, f t represents an output vector, whose dimension values are all between 0 and 1; σ is the sigmoid function; is the i-th column of the weight matrix; is the input vector z t The i-th element of , is the hidden state at the previous moment; x t is the input at the current moment; b f is the bias term of the forget gate; n is the input vector zt Dimensions; The input gate controls how the current input information is updated to the memory unit. The sigmoid function is used to determine the update part, and then the tanh function is used to generate new candidate values, which can be expressed as: ; ; In the formula, i t is the output of the input gate; is the state output of the candidate memory unit; is the jth column of the input gate weight matrix; is the vector z t The jth element of Dimensions; is the kth column of the candidate memory unit state weight matrix; is the vector z t The kth element of b i is the input gate bias; b c Bias the candidate memory cell; The memory unit updates the memory state through the calculation of the forget gate and the input gate, and updates the memory unit, which is expressed as: ; In the formula, C t is the current memory unit state; is the previous memory unit state; W f is the weight matrix of the forget gate; W c is the weight matrix of the candidate memory unit state; W i is the weight matrix of the input gate; The output gate determines the output of the current memory unit, controls the output through the sigmoid function, and transforms the memory unit state through the tanh function, which is expressed as: ; ; In the formula, O t is the output of the output gate; is the jth column of the output gate weight matrix; b o is the bias term of the output gate; W o is the output gate weight matrix; h t is the current hidden state.
[0009] As a preferred solution of the present invention, in S7, the steps of the prediction method based on the deep space analysis engine are specifically as follows: S7.1, standardize and normalize the preprocessed historical data as input data, and set the number of iterations P1; S7.2. Construct a deep spatial analysis engine model. The deep spatial analysis engine model includes three layers of spatial analysis engines, each layer including a forget gate, an input gate, a memory unit, and an output gate; S7.3, perform model training, pass the input data to the first-layer spatial analysis engine, generate the hidden state and unit state of each time step, and calculate the output of each gate. According to the activation value of the output gate, decide which information is output from the current state value to the next layer for calculation. Repeat the above steps until the Q1 group training set data is processed and the training is terminated; S7.4, using the trained deep spatial analysis engine model to predict the wind speed of each wind turbine, the input data passes through each layer in turn, and each layer is passed from the forget gate to the output gate. After passing through the three-layer spatial analysis engine, the prediction result is obtained. The input data is the standardized and normalized data in S71; S7.5. Use the weighted cross entropy loss function to calculate the difference between the model's predicted output and the true value. S7.6. Repeat S7.3-S7.5 until the preset number of iterations P1 is reached or the loss converges, and output the best prediction value of the deep spatial analysis engine.
[0010] As a preferred solution of the present invention, in the above S8, in the time series analysis engine, autoregression means that the current value is associated with the past value, and its order is p, which means that the current value is associated with the value of p periods in the past, and p is determined by the partial autocorrelation function PACF; The matrix form of autoregression is expressed as: ; Where Y is the response variable vector, which contains the p To Y T Time series data, Y T is the current value of the time series, Y p is the pth observation value of the time series; is a coefficient vector containing the coefficients of the autoregressive model to and a constant term a; X is the autoregressive design matrix, each row corresponds to the lagged value and constant term of each observation in the time series, and the first row contains the lagged value and constant term from Y p−1 to the lagged value of Y1, and a constant term 1, and the same applies to the rest of the rows; Y1 is the first observation of the time series, and the rest of the symbols are the same; The overall autoregressive expression is: ; In the formula, is the random error term; The difference eliminates the trend and turns the non-stationary time series into a stationary time series. The order is d, which means that the time series is subjected to a d-order difference operation, which can be expressed as: ; In the formula, is the first-order difference value of the time series at time t; Y t is the tth observation value of the time series; Y t-1 is the t-1th observation value of the time series, that is, the value at the previous time point. The other symbols in the following formula are the same; after d times of difference, it is expressed as: ; In the formula, is the d-order difference value of the time series at time t, corresponding to the difference order d. Starting from d=0, each time the first-order difference is calculated, a stationarity test is performed until the data is stable, which is expressed as: ; The sliding average indicates that the current value is related to the linear combination of the past error terms, whose order is q, indicating that the current value is related to the error terms of the past q periods, and q is determined by the partial autocorrelation function ACF; under the sliding average, Y t It is expressed as: ; In the formula, is the model parameter of the corresponding time series at time t, that is, the random error term, and the other symbols are similar; is corresponding to The sliding average coefficient of , and the other symbols are the same; is the mean, used to provide a benchmark level; The matrix form of the sliding average is expressed as: ; For each random error term, Akaike information criterion AIC and Bayesian information criterion BIC were used to obtain; The timing analysis engine is finally expressed as: .
[0011] As a preferred solution of the present invention, in S9, the steps of the prediction method based on the timing analysis engine are specifically as follows: S9.1. For the time series analysis engine model including autoregression, difference and moving average, define the model parameters: for the difference order d, use the unit root test to determine whether the time series is stable. If it is stable, set d=0. If it is not stable, perform a difference and perform the stability test again; determine p through the partial autocorrelation function PACF graph; determine q through the autocorrelation function ACF graph; S9.2, fitness function definition, using weighted root mean square error W-RMSE and comparative fit index CFI to evaluate the prediction of the time series analysis engine model; S9.3. Calculate the fitness, use the determined values of p, d, and q to build and train the time series analysis engine model, fit it on the Q2 group training set data, calculate the values of the fitness function W-RMSE and CFI, and record the results; S9.4, select the best one, repeat S9.1-S9.3, perform hyperparameter optimization, adjust the values of p and q, compare all model training fitness results, and select the one with the lowest W-RMSE and CFI value greater than or equal to 0.9 as the prediction result; S9.5. The termination condition is that the number of iterations reaches P2 or the training error margin is less than 0.01. When the termination condition is reached, the model training is terminated and the optimal model parameters and the optimal prediction value of the time series analysis engine model are output.
[0012] As a preferred solution of the present invention, in the above S10, the fusion process is to define the best prediction value of the prediction method based on the deep space analysis engine as The best prediction value of the prediction method based on the time series analysis engine is , before fusion, first calculate , The corresponding weighted root mean square errors are expressed as and (Here we can use the well-known W-RMSE calculation formula to obtain it), and then calculate , The corresponding weight , , respectively expressed as: ; ; Perform weighted combination to obtain the final predicted value of wind speed after fusion The expression is: .
[0013] As a preferred solution of the present invention, in the S11, the time-series prediction of the wind speed of the wind turbine is performed using the spatiotemporal intelligent adaptive prediction engine algorithm to obtain the final predicted value of the wind speed. The specific process is to extract the features of the pre-processed historical data and then use the spatiotemporal intelligent adaptive prediction engine algorithm to obtain the final predicted value of the wind speed based on the test set data. and , and then combine the corresponding weights of the two to perform weighted average to obtain the final predicted value of wind speed .
[0014] As a preferred embodiment of the present invention, in S12, the control strategy of the fan is: when the final predicted value of the wind speed is greater than the high wind speed threshold and the closed bus temperature obtained by real-time monitoring is lower than the preset temperature, the output power of the fan is reduced; when the final predicted value of the wind speed is less than the low wind speed threshold and the closed bus temperature obtained by real-time monitoring is higher than the preset temperature, the output power of the fan is increased; in other cases, there is no need to adjust the power output of the fan; integrating the control strategy with the automation system is a configured control system to achieve automatic fan control.
[0015] The algorithm involved in the present invention can be executed by an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the above algorithm calculation is implemented by executing the software by the processor.
[0016] The beneficial effects of the present invention are: High-precision prediction: By using historical data and real-time data, combined with machine learning or deep learning models, the present invention can accurately predict the wind speed in the closed bus. This high-precision prediction helps optimize the operating parameters of the ventilation system based on fans at the closed bus to ensure that the environmental conditions meet specific requirements.
[0017] Dynamic adaptive adjustment: The spatiotemporal intelligent adaptive prediction engine can dynamically adjust the prediction model according to environmental changes and ventilation system status, so that it can more accurately reflect the actual situation. This adaptive capability enables the control system to respond to various disturbances in a timely manner and maintain the stability and efficiency of the ventilation system.
[0018] Improved energy efficiency: By accurately predicting and controlling wind speed, the energy consumption of the ventilation system can be managed more effectively and unnecessary energy waste can be avoided. For example, when lower wind speed is required, the fan speed can be reduced to save energy.
[0019] Prolong equipment life: Reasonable wind speed control can reduce equipment wear and failure rate, because too high or too low wind speed may cause damage to the equipment. Through precise control, the service life of the equipment can be extended and maintenance costs can be reduced.
[0020] Improved working environment: In industrial environments, proper air speed is critical to ensure air quality and worker health. By precisely controlling the air speed within the enclosed busbar, a safer and more comfortable working environment can be created.
[0021] Enhanced system robustness: Spatiotemporal intelligent adaptive prediction engine algorithms are generally more robust and can maintain stable performance in the face of uncertainty and complexity. This means that the ventilation system can operate reliably even when external conditions change.
[0022] Support decision making: Accurate wind speed forecasts are not only beneficial for real-time control, but can also support long-term planning and decision making. For example, when designing a new ventilation system or improving an existing system, these forecast data can be used to optimize the design and operation strategy.
[0023] Easy to integrate and expand: The spatiotemporal adaptive prediction engine algorithm can be seamlessly integrated with other intelligent building management systems to achieve more extensive automation and intelligent control. In addition, as technology develops, the algorithm can be easily upgraded to include new features and improvements. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 It is the comparison between the predicted value and the true value of the spatiotemporal intelligent adaptive prediction engine algorithm in the verification process of the present invention. Figure 1 ; Figure 3 It is the comparison between the predicted value and the true value of the spatiotemporal intelligent adaptive prediction engine algorithm in the verification process of the present invention. Figure 2 ; Figure 4 It is the comparison between the predicted value and the true value of the spatiotemporal intelligent adaptive prediction engine algorithm in the verification process of the present invention. Figure 3 . DETAILED DESCRIPTION
[0025] The embodiments of the present invention are further described below in conjunction with the accompanying drawings: like Figure 1 As shown, the closed bus wind speed prediction control method based on the spatiotemporal intelligent adaptive prediction engine includes the following steps: S1. For the fans at the closed busbar where wind speed prediction is required, wind speed thresholds are set, including high wind speed thresholds and low wind speed thresholds, and a control system is configured. The control system is used to adjust the fan wind speed in combination with the wind speed prediction and the closed busbar temperature obtained by real-time monitoring to ensure that the closed busbar temperature is within a safe range; S2. For wind turbines that need wind speed prediction, collect historical data of wind speed of each wind turbine, process missing values and abnormal values in the historical data, normalize and smooth the processed historical data, complete preprocessing of the historical data, and divide the preprocessed historical data into a training set and a test set; S3. Extract features from the preprocessed historical data based on deep learning and analyze the time series features, which include trends, seasonality, and periodicity. S4. Establish a basic neural network architecture for processing time series features. The basic neural network architecture includes three parts: input layer, hidden layer, and output layer. By connecting the output of the hidden layer to the input of the next time point, a loop structure is formed to capture the time dependency in the time series features. S5. Establish a spatial analysis engine, including a forget gate, an input gate, a memory unit, and an output gate, to optimize the output of the basic neural network architecture; S6. Establish a deep spatial analysis engine. The deep spatial analysis engine consists of multiple layers of spatial analysis engines and can capture more complex data features. S7, constructing a prediction method based on a deep spatial analysis engine, the steps including data preparation, building a deep spatial analysis engine model, model training, wind speed prediction and loss calculation; S8. Establish a time series analysis engine, including autoregression, difference and moving average, for predicting time series data with autoregression and moving average characteristics; S9, constructing a prediction method based on a time series analysis engine, the steps including defining model parameters, defining a fitness function, calculating fitness, selecting the best, and setting a termination condition; S10, integrating the deep spatial analysis engine with the time series analysis engine to obtain a spatiotemporal intelligent adaptive prediction engine algorithm; S11. Apply the spatiotemporal intelligent adaptive prediction engine algorithm to perform time series prediction of wind speed of the wind turbine to obtain the final predicted value of the wind speed; S12. Based on the final predicted value of the wind speed and the closed bus temperature obtained through real-time monitoring, a control strategy for the corresponding fan is formulated, and the fan speed is adjusted through the control system according to the control strategy.
[0026] In S1, the following methods can be used to select wind turbines at closed busbars where wind speed prediction is required: S1.1. Analyze the layout of the enclosed busbar and identify critical areas that may cause insufficient heat dissipation due to space limitations or high current loads. These areas are usually the focus of thermal management. S1.2. Calculate the heat dissipation requirements of each key area based on the specifications and operating conditions of the enclosed busbar (such as current size, ambient temperature, etc.). This includes understanding the performance characteristics of different types of fans, such as air volume, air pressure, noise level, etc., and their performance under different working conditions; S1.3. Based on the above demonstration, select the fan and its model that can meet the heat dissipation requirements, including the fan's size, power, energy efficiency ratio, reliability and maintenance cost. At the same time, ensure the compatibility of the selected fan with the existing system, such as power supply, control interface, etc.
[0027] In this embodiment, the high wind speed threshold and the low wind speed threshold are set to 40 m / s and 27 m / s respectively; when dividing the historical data, it is divided into a training set and a test set in chronological order.
[0028] In S2, wind speed sensors are used to collect historical data of wind speed of each wind turbine (when collecting, ensure that the data covers a sufficient time range to ensure the completeness and accuracy of the data), and interpolation and mean filling are used to supplement missing values; standard scores are used to identify outliers and replace or delete them; When smoothing data, single exponential smoothing is used for data without trend (in actual judgment, we will observe whether there is an obvious trend); double exponential smoothing is used for data with trend (here we mainly look at linear trend) but without seasonality; triple exponential smoothing is used for data with trend and seasonality.
[0029] In S3, for trend analysis, a time series graph of wind speed data is drawn based on the extracted features, and a trend line is fitted to the wind speed time series data using the least squares curve fitting technique to describe the trend in the form of a mathematical model; For seasonal analysis, the extracted features were grouped by season, and the average wind speed for each season was plotted to see if there was an obvious seasonal pattern. The wind speed differences between different seasons were compared to analyze the impact of seasonality on wind speed. For periodicity analysis, the autocorrelation function is calculated and the ACF plot is used to analyze the correlation of the features extracted from wind speed at different time lags to find periodic patterns; check whether the period remains stable over time and whether there is a trend of gradual increase or decrease; In addition, combined with the closed busbar temperature obtained by real-time monitoring, a joint trend analysis of wind speed and temperature characteristics was performed to observe whether temperature changes had a significant impact on wind speed.
[0030] In S5, the forget gate determines how much past information to retain, and calculates the amount of information in the previous state through a sigmoid function: ; In the formula, f t represents an output vector, whose dimension values are all between 0 and 1; σ is the sigmoid function; is the i-th column of the weight matrix; is the input vector z t The i-th element of , is the hidden state at the previous moment; x t is the input at the current moment; b f is the bias term of the forget gate; n is the input vector z t Dimensions; The input gate controls how the current input information is updated to the memory unit. The sigmoid function is used to determine the update part, and then the tanh function is used to generate new candidate values, which can be expressed as: ; ; In the formula, i t is the output of the input gate; is the state output of the candidate memory unit; is the jth column of the input gate weight matrix; is the vector z t The jth element of Dimensions; is the kth column of the candidate memory unit state weight matrix; is the vector z t The kth element of b i is the input gate bias; b c Bias the candidate memory cell; The memory unit updates the memory state through the calculation of the forget gate and the input gate, and updates the memory unit, which is expressed as: ; In the formula, C t is the current memory unit state; is the previous memory unit state; W f is the weight matrix of the forget gate; W c is the weight matrix of the candidate memory unit state; W i is the weight matrix of the input gate; The output gate determines the output of the current memory unit, controls the output through the sigmoid function, and transforms the memory unit state through the tanh function, which is expressed as: ; ; In the formula, O t is the output of the output gate; is the jth column of the output gate weight matrix; b o is the bias term of the output gate; W o is the output gate weight matrix; h t is the current hidden state.
[0031] In S7, the steps of the prediction method based on the deep spatial analysis engine are specifically as follows: S7.1, standardize and normalize the preprocessed historical data as input data, and set the number of iterations P1; S7.2. Construct a deep spatial analysis engine model. The deep spatial analysis engine model includes three layers of spatial analysis engines, each layer including a forget gate, an input gate, a memory unit, and an output gate; S7.3, perform model training, pass the input data to the first-layer spatial analysis engine, generate the hidden state and unit state of each time step, and calculate the output of each gate. According to the activation value of the output gate, decide which information is output from the current state value to the next layer for calculation. Repeat the above steps until the Q1 group training set data is processed and the training is terminated; S7.4, using the trained deep spatial analysis engine model to predict the wind speed of each wind turbine, the input data passes through each layer in turn, and each layer is passed from the forget gate to the output gate. After passing through the three-layer spatial analysis engine, the prediction result is obtained. The input data is the standardized and normalized data in S71; S7.5. Use the weighted cross entropy loss function to calculate the difference between the model's predicted output and the true value. S7.6. Repeat S7.3-S7.5 until the preset number of iterations P1 is reached or the loss converges, and output the best prediction value of the deep spatial analysis engine.
[0032] In S8, in the time series analysis engine, autoregression means that the current value is associated with the past value, and its order is p, which means that the current value is related to the value of p periods in the past, and p is determined by the partial autocorrelation function PACF; The matrix form of autoregression is expressed as: ; Where Y is the response variable vector, which contains the p To Y T Time series data, Y T is the current value of the time series, Y p is the pth observation value of the time series; is a coefficient vector containing the coefficients of the autoregressive model to and a constant term a; X is the autoregressive design matrix, each row corresponds to the lagged value and constant term of each observation in the time series, and the first row contains the lagged value and constant term from Y p−1 to the lagged value of Y1, and a constant term 1, and the same applies to the rest of the rows; Y1 is the first observation of the time series, and the rest of the symbols are the same; The overall autoregressive expression is: ; In the formula, is the random error term; The difference eliminates the trend and turns the non-stationary time series into a stationary time series. The order is d, which means that the time series is subjected to a d-order difference operation, which can be expressed as: ; In the formula, is the first-order difference value of the time series at time t; Y t is the tth observation value of the time series; Y t-1 is the t-1th observation value of the time series, that is, the value at the previous time point. The other symbols in the following formula are the same; after d times of difference, it is expressed as: ; In the formula, is the d-order difference value of the time series at time t, corresponding to the difference order d. Starting from d=0, each time the first-order difference is calculated, a stationarity test is performed until the data is stable, which is expressed as: ; The sliding average indicates that the current value is related to the linear combination of the past error terms, whose order is q, indicating that the current value is related to the error terms of the past q periods, and q is determined by the partial autocorrelation function ACF; under the sliding average, Y t It is expressed as: ; In the formula, is the model parameter of the corresponding time series at time t, that is, the random error term, and the other symbols are similar; is corresponding to The sliding average coefficient of , and the other symbols are the same; is the mean, used to provide a benchmark level; The matrix form of the sliding average is expressed as: ; For each random error term, Akaike information criterion AIC and Bayesian information criterion BIC were used to obtain; The timing analysis engine is finally expressed as: .
[0033] In S9, the steps of the prediction method based on the time series analysis engine are specifically as follows: S9.1. For the time series analysis engine model including autoregression, difference and moving average, define the model parameters: for the difference order d, use the unit root test to determine whether the time series is stable. If it is stable, set d=0. If it is not stable, perform a difference and perform the stability test again; determine p through the partial autocorrelation function PACF graph; determine q through the autocorrelation function ACF graph; S9.2, fitness function definition, using weighted root mean square error W-RMSE and comparative fit index CFI to evaluate the prediction of the time series analysis engine model; S9.3. Calculate the fitness, use the determined values of p, d, and q to build and train the time series analysis engine model, fit it on the Q2 group training set data, calculate the values of the fitness function W-RMSE and CFI, and record the results; S9.4, select the best one, repeat S9.1-S9.3, perform hyperparameter optimization, adjust the values of p and q, compare all model training fitness results, and select the one with the lowest W-RMSE and CFI value greater than or equal to 0.9 as the prediction result; S9.5. The termination condition is that the number of iterations reaches P2 or the training error margin is less than 0.01. When the termination condition is reached, the model training is terminated and the optimal model parameters and the optimal prediction value of the time series analysis engine model are output.
[0034] The candidate values of p and q are identified by the autocorrelation function graph and the partial correlation function graph. If the PACF graph is quickly truncated after a certain lag p, it usually indicates that p is the lag value; if the ACF graph is quickly truncated after a certain lag q, it usually indicates that q is the lag value.
[0035] In this embodiment, the values of P1 and P1 are 200.
[0036] In S10, the fusion process is to define the best prediction value of the prediction method based on the deep spatial analysis engine as The best prediction value of the prediction method based on the time series analysis engine is , before fusion, first calculate , The corresponding weighted root mean square errors are expressed as and , then calculate , The corresponding weight , , respectively expressed as: ; ; Perform weighted combination to obtain the final predicted value of wind speed after fusion The expression is: .
[0037] In S11, the spatiotemporal intelligent adaptive prediction engine algorithm is used to perform time series prediction of wind turbine wind speed to obtain the final predicted value of wind speed. The specific process is to extract features from the preprocessed historical data and use the spatiotemporal intelligent adaptive prediction engine algorithm to obtain the final predicted value of wind speed based on the test set data and the test set data, respectively. and , and then combine the corresponding weights of the two to perform weighted average to obtain the final predicted value of wind speed .
[0038] In S12, the control strategy of the fan is to reduce the output power of the fan when the final predicted value of the wind speed is greater than the high wind speed threshold and the closed bus temperature obtained by real-time monitoring is lower than the preset temperature; to increase the output power of the fan when the final predicted value of the wind speed is less than the low wind speed threshold and the closed bus temperature obtained by real-time monitoring is higher than the preset temperature; in other cases, there is no need to adjust the power output of the fan; integrating this control strategy with the automation system is a configured control system to achieve automatic fan control.
[0039] Furthermore, it may also include: S13. When dividing the preprocessed historical data, an additional validation set is divided, and the validation set data is used as the basis for evaluation. The prediction value of the spatiotemporal intelligent adaptive prediction engine algorithm is evaluated using the evaluation function. Specifically, the weighted mean square error (W-RMSE) can be used to evaluate the prediction value. The lower the W-RMSE value, the higher the prediction accuracy. The comparative fit index (CFI) of the prediction value is calculated. The CFI result is usually between 0 and 1. The closer the CFI value is to 1, the stronger the model's fitting ability. According to the evaluation results, analyze the shortcomings of the algorithm performance, consider adjusting the model parameters, weights, or using more complex integration methods. If necessary, retrain and optimize the model after collecting new data to improve the prediction accuracy.
[0040] S14. Calculate the error based on the wind speed data collected in real time by the wind speed sensor and the wind speed data predicted by the prediction model (time-space intelligent adaptive prediction engine algorithm), analyze the prediction error, and identify the reasons for inaccurate model prediction (including seasonal changes and meteorological conditions). Regularly use the latest actual wind speed data, retrain the prediction model, update the control strategy, and form a closed-loop feedback process, which can effectively implement a closed-loop control system based on actual wind speed feedback, continuously optimize wind speed prediction and wind turbine control strategy, and improve power generation efficiency and operational safety.
[0041] Based on the prediction method of the present invention, a prediction is made (a predicted value is obtained), and the predicted value is compared with the wind speed data (real value) collected in real time. The comparison result is as follows: Figure 2-Figure 4 As shown, Figure 2 This is the comparison when the number of samples in the training data is 8 groups. Figure 3 This is the comparison when the number of samples in the training data is 50 groups. Figure 4 This is the comparison when the number of samples in the training data is 200 groups.
[0042] The comparison results show that most of the predicted values are close to the true values, especially in the relatively stable wind speed range, indicating that the prediction method of the present invention has high accuracy in the relatively stable wind speed range. In some extreme fluctuation positions (such as when the wind speed suddenly rises or falls), there is a deviation between the predicted value and the true value, which may be due to the limitations of the present invention in capturing the rapid changes in wind speed, but this situation is rare. In general, the present invention performs well in capturing the overall trend and regular fluctuations of wind speed, and can be applied to a variety of wind speed prediction tasks.
Claims
1. A closed busbar wind speed prediction and control method based on a spatiotemporal intelligent adaptive prediction engine is characterized by: The following steps are involved: S1. For the fans at the closed busbar where wind speed prediction is required, wind speed thresholds are set, including high wind speed thresholds and low wind speed thresholds, and a control system is configured. The control system is used to adjust the fan wind speed in combination with the wind speed prediction and the closed busbar temperature obtained by real-time monitoring to ensure that the closed busbar temperature is within a safe range; S2. For wind turbines that need wind speed prediction, collect historical data of wind speed of each wind turbine, process missing values and abnormal values in the historical data, normalize and smooth the processed historical data, complete preprocessing of the historical data, and divide the preprocessed historical data into a training set and a test set; S3. Extract features from the preprocessed historical data based on deep learning and analyze the time series features, which include trends, seasonality, and periodicity. S4. Establish a basic neural network architecture for processing time series features. The basic neural network architecture includes three parts: input layer, hidden layer, and output layer. By connecting the output of the hidden layer to the input of the next time point, a loop structure is formed to capture the time dependency in the time series features. S5. Establish a spatial analysis engine, including a forget gate, an input gate, a memory unit, and an output gate, to optimize the output of the basic neural network architecture; S6. Establish a deep spatial analysis engine. The deep spatial analysis engine consists of multiple layers of spatial analysis engines and can capture more complex data features. S7, constructing a prediction method based on a deep spatial analysis engine, the steps including data preparation, building a deep spatial analysis engine model, model training, wind speed prediction and loss calculation; S8. Establish a time series analysis engine, including autoregression, difference and moving average, for predicting time series data with autoregression and moving average characteristics; S9, constructing a prediction method based on a time series analysis engine, the steps including defining model parameters, defining a fitness function, calculating fitness, selecting the best, and setting a termination condition; S10, integrating the deep spatial analysis engine with the time series analysis engine to obtain a spatiotemporal intelligent adaptive prediction engine algorithm; S11. Apply the spatiotemporal intelligent adaptive prediction engine algorithm to perform time series prediction of wind speed of the wind turbine to obtain the final predicted value of the wind speed; S12. Based on the final predicted value of the wind speed and the closed bus temperature obtained through real-time monitoring, a control strategy for the corresponding fan is formulated, and the fan speed is adjusted through the control system according to the control strategy.
2. According to claim 1, the closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine is characterized by: In S2, the wind speed sensor is used to collect the historical data of the wind speed of each wind turbine, and the missing values are supplemented by interpolation and mean filling; the outliers are identified by using standard scores, and replaced or deleted; When smoothing data, single exponential smoothing is used for data without trend; double exponential smoothing is used for data with trend but no seasonality; and triple exponential smoothing is used for data with trend and seasonality.
3. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 1 is characterized in that: In the above S3, for trend analysis, a time series diagram of wind speed data is drawn based on the extracted features, and a trend line is fitted to the wind speed time series data using a least squares curve fitting technique to describe the trend in the form of a mathematical model; For seasonal analysis, the extracted features were grouped by season, and the average wind speed for each season was plotted to see if there was an obvious seasonal pattern. The wind speed differences between different seasons were compared to analyze the impact of seasonality on wind speed. For periodicity analysis, the autocorrelation function is calculated and the ACF plot is used to analyze the correlation of the features extracted from wind speed at different time lags to find periodic patterns; check whether the period remains stable over time and whether there is a trend of gradual increase or decrease; In addition, combined with the closed busbar temperature obtained by real-time monitoring, a joint trend analysis of wind speed and temperature characteristics was performed to observe whether temperature changes had a significant impact on wind speed.
4. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 1 is characterized in that: In S5, the forget gate determines how much past information to retain, and calculates the amount of information in the previous state through a sigmoid function: ; In the formula, f t represents an output vector, whose dimension values are all between 0 and 1; σ is the sigmoid function; is the i-th column of the weight matrix; is the input vector z t The i-th element of , is the hidden state at the previous moment; x t is the input at the current moment; b f is the bias term of the forget gate; n is the input vector z t Dimensions; The input gate controls how the current input information is updated to the memory unit. The sigmoid function is used to determine the update part, and then the tanh function is used to generate new candidate values, which can be expressed as: ; ; In the formula, i t is the output of the input gate; is the state output of the candidate memory unit; is the jth column of the input gate weight matrix; is the vector z t The jth element of Dimensions; is the kth column of the candidate memory unit state weight matrix; is the vector z t The kth element of b i is the input gate bias; b c Bias the candidate memory cell; The memory unit updates the memory state through the calculation of the forget gate and the input gate, and updates the memory unit, which is expressed as: ; In the formula, C t is the current memory unit state; is the previous memory unit state; W f is the weight matrix of the forget gate; W c is the weight matrix of the candidate memory unit state; W i is the weight matrix of the input gate; The output gate determines the output of the current memory unit, controls the output through the sigmoid function, and transforms the memory unit state through the tanh function, which is expressed as: ; ; In the formula, O t is the output of the output gate; is the jth column of the output gate weight matrix; b o is the bias term of the output gate; W o is the output gate weight matrix; h t is the current hidden state.
5. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 1 is characterized in that: In S7, the steps of the prediction method based on the deep space analysis engine are specifically as follows: S7.1, standardize and normalize the preprocessed historical data as input data, and set the number of iterations P1; S7.
2. Construct a deep spatial analysis engine model. The deep spatial analysis engine model includes three layers of spatial analysis engines, each layer including a forget gate, an input gate, a memory unit, and an output gate; S7.3, perform model training, pass the input data to the first-layer spatial analysis engine, generate the hidden state and unit state of each time step, and calculate the output of each gate. According to the activation value of the output gate, decide which information is output from the current state value to the next layer for calculation. Repeat the above steps until the Q1 group training set data is processed and the training is terminated; S7.4, using the trained deep spatial analysis engine model to predict the wind speed of each wind turbine, the input data passes through each layer in turn, and each layer is passed from the forget gate to the output gate. After passing through the three-layer spatial analysis engine, the prediction result is obtained. The input data is the standardized and normalized data in S71; S7.
5. Use the weighted cross entropy loss function to calculate the difference between the model's predicted output and the true value. S7.
6. Repeat S7.3-S7.5 until the preset number of iterations P1 is reached or the loss converges, and output the best prediction value of the deep spatial analysis engine.
6. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 1 is characterized in that: In the above S8, in the time series analysis engine, autoregression means that the current value is associated with the past value, and its order is p, which means that the current value is associated with the value of p periods in the past, and p is determined by the partial autocorrelation function PACF; The matrix form of autoregression is expressed as: ; Where Y is the response variable vector, which contains the p To Y T Time series data, Y T is the current value of the time series, Y p is the pth observation value of the time series; is a coefficient vector containing the coefficients of the autoregressive model to and a constant term a; X is the autoregressive design matrix, each row corresponds to the lagged value and constant term of each observation in the time series, and the first row contains the lagged value and constant term from Y p−1 The lag value to Y1, and a constant term 1, and the rest of the rows are the same; Y1 is the first observation of the time series, and the rest of the symbols are the same; The overall autoregressive expression is: ; In the formula, is the random error term; The difference eliminates the trend and turns the non-stationary time series into a stationary time series. The order is d, which means that the time series is subjected to a d-order difference operation, which can be expressed as: ; In the formula, is the first-order difference value of the time series at time t; Y t is the tth observation value of the time series; Y t-1 is the t-1th observation value of the time series, that is, the value at the previous time point. The other symbols in the following formula are the same; after d times of difference, it is expressed as: ; In the formula, is the d-order difference value of the time series at time t, corresponding to the difference order d. Starting from d=0, each time the first-order difference is calculated, a stationarity test is performed until the data is stable, which is expressed as: ; The sliding average indicates that the current value is related to the linear combination of the past error terms, whose order is q, indicating that the current value is related to the error terms of the past q periods, and q is determined by the partial autocorrelation function ACF; under the sliding average, Y t It is expressed as: ; In the formula, is the model parameter of the corresponding time series at time t, that is, the random error term, and the other symbols are similar; is corresponding to The sliding average coefficient of , and the other symbols are the same; is the mean, used to provide a benchmark level; The matrix form of the sliding average is expressed as: ; For each random error term, Akaike information criterion AIC and Bayesian information criterion BIC were used to obtain; The timing analysis engine is finally expressed as: 。 7. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 6 is characterized in that: In the above S9, the steps of the prediction method based on the time series analysis engine are specifically as follows: S9.
1. For the time series analysis engine model including autoregression, difference and moving average, define the model parameters: for the difference order d, use the unit root test to determine whether the time series is stable. If it is stable, set d=0. If it is not stable, perform a difference and perform the stability test again; determine p through the partial autocorrelation function PACF graph; determine q through the autocorrelation function ACF graph; S9.2, fitness function definition, using weighted root mean square error W-RMSE and comparative fit index CFI to evaluate the prediction of the time series analysis engine model; S9.
3. Calculate the fitness, use the determined values of p, d, and q to build and train the time series analysis engine model, fit it on the Q2 group training set data, calculate the values of the fitness function W-RMSE and CFI, and record the results; S9.4, select the best one, repeat S9.1-S9.3, perform hyperparameter optimization, adjust the values of p and q, compare all model training fitness results, and select the one with the lowest W-RMSE and CFI value greater than or equal to 0.9 as the prediction result; S9.
5. The termination condition is that the number of iterations reaches P2 or the training error margin is less than 0.
01. When the termination condition is reached, the model training is terminated and the optimal model parameters and the optimal prediction value of the time series analysis engine model are output.
8. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 1 is characterized in that: In the above S10, the fusion process is to define the best prediction value of the prediction method based on the deep space analysis engine as The best prediction value of the prediction method based on the time series analysis engine is , before fusion, first calculate , The corresponding weighted root mean square errors are expressed as and , then calculate , The corresponding weight , , respectively expressed as: ; ; Perform weighted combination to obtain the final predicted value of wind speed after fusion The expression is: 。 9. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 8 is characterized in that: In the above S11, the time-series prediction of wind speed of the wind turbine is performed by applying the spatiotemporal intelligent adaptive prediction engine algorithm to obtain the final predicted value of wind speed. The specific process is to extract features from the pre-processed historical data and then use the spatiotemporal intelligent adaptive prediction engine algorithm to obtain the final predicted value of wind speed based on the test set data. and , and then combine the corresponding weights of the two to perform weighted average to obtain the final predicted value of wind speed .
10. The closed bus wind speed prediction and control method based on the spatiotemporal intelligent adaptive prediction engine according to claim 1 is characterized in that: In the S12 described above, the control strategy of the fan is to reduce the output power of the fan when the final predicted value of the wind speed is greater than the high wind speed threshold and the closed bus temperature obtained by real-time monitoring is lower than the preset temperature; increase the output power of the fan when the final predicted value of the wind speed is less than the low wind speed threshold and the closed bus temperature obtained by real-time monitoring is higher than the preset temperature; in other cases, there is no need to adjust the power output of the fan; integrating the control strategy with the automation system is a configured control system to achieve automatic fan control.
Citation Information
Patent Citations
Wind speed prediction method and system based on long-term and short-term memory time neural network
CN111222677A
Data center load prediction method based on ensemble learning
CN117272240A
Photovoltaic power generation system management platform and method based on digital twinning and deep learning
CN118572893A
Prediction method for aero-engine starting exhaust temperature
US20200234165A1
