A method and system for improving time series prediction effect

Abstract

The application discloses a method for improving time series prediction effect, integrates and accumulates a certain length of historical time series, or integrates and accumulates after a certain algebraic treatment, and is called a historical integral sequence; a group of fixed length historical integral sequences are used as prediction model inputs, and a sequence of one or more time points is output after calculation of the prediction model, and is called a future integral sequence prediction value; the future integral sequence is subjected to difference calculation to obtain a future time series prediction value. The application is used for realizing faster modeling and more accurate prediction in a time series prediction process. In combination with an integral difference algorithm and a deep learning model, the method reduces model complexity caused by multiple sub-sequences, has clear physical meaning, and improves the accuracy of time series prediction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence time series prediction, and specifically relates to a method and system for improving the performance of time series prediction. Background Technology

[0002] With the development of artificial intelligence and time series forecasting methods, time series information has been widely used in various fields. Utilizing historical data to predict future data trends and guide decision-making is of great significance to industry. However, for random, highly fluctuating time series data (such as wind speed, solar radiation, power load, and financial data), existing deep learning-based signal decomposition methods have unsatisfactory prediction results. The main reasons are that these methods rely on subjective experience, subsequences lack clear physical meaning, high-frequency subsequences are difficult to predict, and they are highly sensitive to new data input.

[0003] To improve modeling efficiency and prediction accuracy, this invention innovatively proposes an "integration-modeling-difference" prediction method, which reduces the model complexity caused by multiple subsequences, has clear physical meaning, and improves the accuracy of time series prediction. Summary of the Invention

[0004] In order to solve the technical problems existing in the background art, the present invention aims to provide a method and system for improving the performance of time series forecasting, so as to achieve faster modeling and more accurate forecasting in the process of time series forecasting.

[0005] To solve the technical problem, the technical solution of the present invention is as follows:

[0006] A method for improving time series forecasting performance includes an integral preprocessing sub-step, a forecasting model, and a difference backtracking target variable sub-step. The time series data described here is a sequence of data points arranged in chronological order, with each data point being a real value.

[0007] The basic process of this prediction method is as follows: integrate and accumulate a historical time series of a certain length, or integrate and accumulate after certain algebraic processing, which is called the historical integral sequence; take a set of historical integral sequences of a fixed length as the input of the prediction model, and output the sequence of one or more time points after the prediction model is calculated, which is called the future integral sequence prediction value; perform difference calculation on the future integral sequence to obtain the future time series prediction value.

[0008] The goal of a forecasting model is to minimize the difference between the predicted and actual values ​​of future time series data. To achieve this, the time series forecasting model needs to be trained using existing time series data, i.e., past data, to obtain its parameters. This time series forecasting model can then be used to predict future time series.

[0009] Since the prediction model of the present invention can only be used after training, the specific operation steps of the present invention are divided into: a prediction model training step and a prediction model testing step. Among them, the training step is to learn all parameters in the time series prediction model; the testing step is to use the time series prediction model to output future data. The specific steps are as follows:

[0010] Step 1: Collect time series data of a certain length, and divide it into a training set and a test set (note: any time series used in the following Steps 2 to 4 comes from the training set);

[0011] Step 2: Perform integration processing on the historical time series to obtain a historical integration series, and normalize the historical integration series;

[0012] Step 3: Initialize the prediction model, and input the historical integration series of the training set into the model for training;

[0013] Step 4: Repeatedly train the prediction model through the grid search algorithm, and select the optimal hyperparameters of the prediction model;

[0014] Step 5: After normalizing the historical integration series of the test set, input it into the trained model to obtain the predicted value of the future integration series of the test set;

[0015] Step 6: Perform denormalization and difference operations on the predicted value of the future integration series of the test set to obtain the predicted value of the future time series;

[0016] Step 7: Evaluate the prediction accuracy on the test set.

[0017] In Step 1, dividing the training set and the test set: Divide the entire data set into a training set and a test set in a ratio of 70% / 30% in chronological order, that is, the first 70% is the training set, and the last 30% is the test set. Suppose the training set contains a total of N time series data. For any time series data (y1, y2,..., y t ), select the parameter l (l < t). The subsequence composed of the first l data of the sequence is the historical time series, and the subsequence composed of the data from the (l + 1)-th to the (l + a)-th data is the future time series, denoted as (y l+1 , y l+2 ,..., y l+a ).

[0018] The integration processing described in Step 2 is defined as that the integrated quantity Y i at a certain time stamp is equal to the accumulation of all time series data from the first time series data Y1 to the time series Y t corresponding to this time stamp, that is Thus, the integrated time series (Y1, Y2,..., YN Integration can transform a fluctuating sequence into a smooth, monotonically increasing sequence. The normalization method uses linear normalization, i.e. This allows the numerical range of the integral sequence to be controlled within [0,1], which facilitates subsequent calculations.

[0019] The prediction model described in step 3 can be any existing time series prediction model, such as Support Vector Regression (SVR), Backpropagation Neural Network (BPNN), Long Short-Term Memory Neural Network (LSTM), and Gated Recurrent Unit Neural Network (GRU). In this invention, considering both prediction accuracy and parameter computation, GRU is ultimately selected as the prediction model. This model consists of one GRU layer and one fully connected layer. The computational unit of the GRU layer is a multi-layer fully connected neural network with ReLU as the activation function. At each time node t, the input layer dimension of this recurrent neural network is... The output is The output represents the prediction for time point t+1 based on time point t. The fully connected layer uses ReLU as the activation function, and its input is the output of the GRU layer. Its output is a one-dimensional vector of length a, which is the predicted value of the future integral sequence.

[0020] In step 4, the grid search algorithm, also known as exhaustive search, iterates through all candidate hyperparameter combinations and selects the best-performing parameter as the final result.

[0021] In step 6, inverse normalization is the inverse operation of normalization, i.e., Y i =Y′ i (Y max -Y min )+Y min As defined in step 2, the value of the time series corresponding to a certain timestamp is equal to the first difference between the integrated time series values ​​of the current timestamp and the previous timestamp, i.e., y. t =Y t -Y t-1 Therefore, future time series forecasts can be obtained.

[0022] Step 7 uses Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) to evaluate the accuracy of time series forecasts. Smaller MAE and RMSE indicate higher forecast accuracy. For the actual future time series value y... i and their corresponding predicted values MAE and RMSE can be represented as follows:

[0023]

[0024]

[0025] The present invention also provides a system for improving the performance of time series forecasting, comprising:

[0026] SCADA systems are used for data acquisition and transmission;

[0027] One or more processors;

[0028] Memory, used to store algorithm programs and data transmitted by the SCADA system;

[0029] The SCADA system collects field data and transmits it to a memory for storage. When the algorithm program is executed by the one or more processors, the one or more processors execute the aforementioned method for improving time series prediction performance.

[0030] Compared with the prior art, the advantages of the present invention are as follows:

[0031] (1) The present invention uses variable upper limit integral to smooth the high frequency fluctuations of time series, making it easier for machine learning models to learn the characteristics of smooth curves.

[0032] (2) This invention avoids decomposing time series and modeling them separately, which greatly reduces the number of sub-models, reduces model complexity, and improves computational efficiency.

[0033] (3) The integral time series of the present invention still has a clear physical meaning and has stronger interpretability compared with the subsequence of the decomposition model. Attached Figure Description

[0034] Figure 1 This is a flowchart of the algorithm of the present invention;

[0035] Figure 2 This is a comparison chart of the prediction results of the time prediction system of this invention with those of EMD and Persistence methods.

[0036] Figure 3 This is a comparison chart of the model training time of the time prediction system of this invention and the EMD method. Detailed Implementation

[0037] The specific technical solutions of the present invention will be described with reference to the embodiments.

[0038] This implementation uses the GEFCom2014 wind power dataset and improves the prediction accuracy of the time series forecasting system at future time points using the method proposed in this invention. The specific steps of this embodiment are as follows: Figure 1 As shown, it includes:

[0039] Step 1: Collect time series data of a certain length and divide it into training and test sets;

[0040] Step 2: Integrate the historical time series to obtain a historical integrated series, and normalize the historical integrated series;

[0041] Step 3: Initialize the prediction model, and input the historical integrated series of the training set into the model for training;

[0042] Step 4: Repeatedly train the prediction model through the grid search algorithm, and select the optimal hyperparameters of the prediction model;

[0043] Step 5: After normalizing the historical integrated series of the test set, input it into the trained model to obtain the predicted values of the future integrated series of the test set;

[0044] Step 6: Denormalize and perform difference operations on the predicted values of the future integrated series of the test set to obtain the predicted values of the future time series;

[0045] Step 7: Evaluate the prediction accuracy on the test set.

[0046] In Step 1 for dividing the training set and the test set: Since the data set contains wind power data for two years from 2012 to 2013, the data in 2012 is used as the training set, and the data in 2013 is used as the test set. The training set contains a total of 8770 time series data. For any time series data (y1, y2, …, y t ), given the parameter l (l < t), the subsequence composed of the first l data of the sequence is the historical time series, and the subsequence composed of the data from the (l + 1)-th to the (l + a)-th data is the future time series, denoted as (y l+1 , y l+2 , …, y l+a ).

[0047] The integration process described in Step 2 is defined as that the integrated quantity Y i at a certain time stamp is equal to the accumulation of all time series data from the first time series data Y1 to the time series Y t corresponding to this time stamp, that is From this, the integrated time series (Y1, Y2, …, Y N ) can be obtained. The integration process can transform the fluctuating sequence into a smooth and monotonically increasing sequence. The normalization method uses linear normalization, that is From this, the numerical range of the integrated series can be controlled within [0, 1], which is convenient for subsequent calculations.

[0048] The prediction model described in Step 3 uses a gated recurrent unit neural network (GRU). This model contains one GRU layer and one fully connected layer. The computing unit of the GRU layer is a multi-layer fully connected neural network with the ReLU activation function. At each time node t, the input layer dimension of this recurrent neural network is The output is The output represents the prediction for time point t+1 based on time point t. The fully connected layer uses ReLU as the activation function, and its input is the output of the GRU layer. Its output is a one-dimensional vector of length a, which is the predicted value of the future integral sequence.

[0049] The grid search algorithm in step 4 is an exhaustive search. It iterates through all candidate hyperparameter combinations and selects the best-performing parameter as the final result.

[0050] Step 6, inverse normalization, is the inverse operation of normalization, i.e., Y i =Y′ i (Y max -Y min )+Y min As defined in step 2, the value of the time series corresponding to a certain timestamp is equal to the first difference between the integrated time series values ​​of the current timestamp and the previous timestamp, i.e., y. t =Y t -Y t-1 Therefore, future time series forecasts can be obtained.

[0051] In step 7 of this invention, the mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the accuracy of time series prediction. The smaller the MAE and RMSE, the higher the prediction accuracy. Taking a two-step advance prediction as an example, for the actual value y of the future time series... i and their corresponding predicted values MAE and RMSE can be represented as follows:

[0052]

[0053]

[0054] In this example, MAE = 0.06 and RMSE = 0.11.

[0055] Figure 2 This is a comparison chart of the prediction results of the time prediction system of this invention with those of EMD and Persistence methods.

[0056] Figure 3 This is a comparison chart of the model training time of the time prediction system of this invention and the EMD method.

[0057] A system for improving time series forecasting performance includes:

[0058] SCADA systems are used for data acquisition and transmission;

[0059] One or more processors;

[0060] Memory, used to store algorithm programs and data transmitted by the SCADA system;

[0061] The SCADA system collects field data and transmits it to a memory for storage. When the algorithm program is executed by the one or more processors, the one or more processors execute the method described above for improving the time series prediction effect.

[0062] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention. Many other changes and modifications can be made without departing from the concept and scope of the present invention. It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. A method for improving the performance of time series forecasting, characterized in that, Includes the following steps: The integration and summation of historical time series of a certain length, or the integration and summation after certain algebraic processing, is called the historical integral sequence; the prediction model takes a set of historical integral sequences of a fixed length as input, and outputs the sequence of one or more time points after calculation by the prediction model, which is called the future integral sequence prediction value; the future integral sequence prediction value is obtained by performing difference calculation on the future integral sequence. The goal of the prediction model is to minimize the difference between the predicted value and the actual value of the future time series. The prediction model needs to be trained on existing time series data, i.e., past data, to obtain the parameters of the prediction model. This prediction model is used to predict future time series. The specific steps are as follows: Step 1: Collect time series data of a certain length, and divide it into training set and test set. Any time series used in steps 2 to 4 comes from the training set; the time series data is wind power data. Step 2: Integrate the historical time series to obtain the historical integral series, and then normalize the historical integral series; Integration processing is defined as the amount of integration Y at a certain timestamp. i It equals the time series from the first time series data Y1 to the time series Y corresponding to that timestamp. t The summation of all time series data corresponding to the data, i.e. From this, we can obtain the integral time series (Y1, Y2, ..., Y). N Integration can transform a fluctuating sequence into a smooth, monotonically increasing sequence; the normalization method uses linear normalization, i.e. This controls the numerical range of the integral sequence to [0,1], which facilitates subsequent calculations; Step 3: Initialize the prediction model by inputting the historical integral sequences from the training set into the model for training; The prediction model employs a gated recurrent unit neural network (GRU). This model consists of one GRU layer and one fully connected layer. The computational unit of the GRU layer is a multi-layer fully connected neural network with ReLU activation function. At each time point t, the input layer dimension of this recurrent neural network is... The output is The output represents the prediction of the (t+1)th time point based on the tth time point; the fully connected layer uses ReLU as the activation function, its input is the output of the GRU layer, and its output is a one-dimensional vector of length a, which is the predicted value of the future integral sequence; Step 4: Repeatedly train the prediction model using the grid search algorithm and select the optimal hyperparameters of the prediction model; Step 5: After normalizing the historical integral sequence of the test set, input it into the trained model to obtain the predicted value of the future integral sequence of the test set; Step 6: Perform inverse normalization and differencing operations on the predicted future integral sequence values ​​of the test set to obtain the predicted future time series values; Step 7: Evaluate the prediction accuracy on the test set.

2. The method for improving time series forecasting performance according to claim 1, characterized in that, The specific method for dividing the training set and the test set in Step 1 is as follows: The entire data set is divided into a training set and a test set in a 70% / 30% ratio in chronological order, that is, the first 70% is the training set and the last 30% is the test set. Suppose the training set contains a total of N time series data. For any time series data (y1, y2, …, y t ), given the parameter l, where l < t, the subsequence composed of the first l data of the sequence is the historical time series, and the subsequence composed of the data from the (l + 1)-th to the (l + a)-th is the future time series, denoted as (y l+1 , y l+2 , …, y l+a ).

3. The method for improving time series forecasting performance according to claim 1, characterized in that, In step 4, the grid search algorithm, also known as exhaustive search, iterates through all candidate hyperparameter combinations and selects the best-performing parameter as the final result.

4. The method for improving time series forecasting performance according to claim 3, characterized in that, In step 6, inverse normalization is the inverse operation of normalization, i.e., Y i =Y′ i (Y max -Y min )+Y min As defined in step 2, the value of the time series corresponding to a certain timestamp is equal to the first difference between the integrated time series values ​​of the current timestamp and the previous timestamp, i.e., y t =Y t -Y t-1 Therefore, the predicted values ​​for future time series can be obtained.

5. The method for improving time series forecasting performance according to claim 4, characterized in that, Step 7 uses the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) to evaluate the time series forecast accuracy. Smaller MAE and RMSE indicate higher forecast accuracy. For the actual future time series value y... i and their corresponding predicted values MAE and RMSE are respectively represented as 6. A system for improving time series forecasting performance, characterized in that, include: SCADA systems are used for data acquisition and transmission; One or more processors; Memory, used to store algorithm programs and data transmitted by the SCADA system; The SCADA system collects field data and transmits it to a memory for storage. When the algorithm program is executed by the one or more processors, the one or more processors execute the method for improving time series prediction as described in any one of claims 1 to 5.

Images