Time sequence prediction method, system and equipment based on multi-objective design and medium

Abstract

The invention discloses a time sequence prediction method, system and device based on multi-objective design and a medium, the method comprises the steps that time sequence historical data is collected, the time sequence historical data comprises data points xt of a plurality of time points, and t represents a time index; the time sequence historical data is converted into a two-dimensional training data set, the two-dimensional training data set comprises a feature matrix F and a multi-target vector T, the dimension of the feature matrix F is m * n and represents m samples and n features, and the dimension of the multi-target vector T is m * k and represents k target columns; based on the two-dimensional training data set, a tree regression model is used for training and prediction, and a prediction time sequence value is generated and output; and outputting a feature importance sequence of the tree regression model for decision support. According to the method, low error propagation, low calculation complexity and high interpretability of multi-step prediction are considered in time sequence prediction.

Description

Technical Field

[0001] The present invention belongs to the field of big data processing technology, and in particular relates to a time series prediction method, system, device and medium based on multi-objective design. Background Art

[0002] Time series forecasting scenarios (such as water, electricity, and gas loads in energy consumption forecasting) are widespread in smart factory projects. Current mainstream solutions rely on traditional time series models and deep learning methods. Statistical models such as ARIMA and Prophet are often used in single-dimensional forecasting scenarios. Their advantage lies in their high interpretability, but they strictly rely on the assumption of data stationarity and have difficulty handling nonlinear trends or external feature interference. For multi-dimensional scenarios, deep learning methods (such as LSTM and Seq2seq) are used. They can capture long-term temporal dependencies, but they have significant drawbacks: these models require massive amounts of training data and complex computing resources (such as GPU acceleration), resulting in high training costs and inefficient deployment. Furthermore, deep learning models are essentially "black box" structures, lacking interpretability and inability to provide feature importance analysis, limiting their application in decision support.

[0003] Furthermore, existing technologies face the problem of a limited range of algorithmic choices. Traditional methods are primarily limited to time series-specific models (such as ARIMA and LSTM). While tree-based regression models (such as XGBoost and LightGBM), while offering advantages in efficiency and interpretability, are difficult to apply directly to time series data. This is because time series data typically exists in sequence form, while tree models require input as two-dimensional structured datasets (such as feature matrices and multi-target vectors). Current practice requires manual construction of lagged features or sliding window statistics, which increases the complexity of feature engineering. Furthermore, for multi-step forecasting, a separate model is typically trained for each target, leading to error accumulation (error propagation from one level to another in recursive forecasting) and computational redundancy, ultimately impacting forecast accuracy and practicality.

[0004] In summary, the core challenge of existing technologies is how to balance low error propagation, low computational complexity, and high interpretability for multi-step time series forecasting. Specifically, recursive forecasting methods (such as ARIMA) inevitably suffer from error accumulation; deep learning models are computationally expensive and uninterpretable; and the application of tree-based models is limited by bottlenecks in time series data conversion, making it impossible to directly achieve simultaneous multi-target forecasting. This shortcoming is particularly prominent in smart factory scenarios, and an innovative solution is urgently needed to unify these requirements. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies in the above-mentioned prior art and provide a time series prediction method, system, device and medium based on multi-objective design, which takes into account low error propagation, low computational complexity and high interpretability of multi-step prediction in time series prediction.

[0006] The first aspect of the present invention discloses a time series prediction method based on multi-objective design, comprising the following steps: S1: Collecting time series historical data, the time series historical data includes data points x at multiple time points t , where t represents the time index; S2: Convert the time series historical data into a two-dimensional training data set, which includes a feature matrix F and a multi-target vector T, wherein the dimension of the feature matrix F is m×n, representing m samples and n features, and the dimension of the multi-target vector T is m×k, representing k target columns; wherein, x t is the data point at time t, m is the number of samples, n is the feature dimension, and k is the number of targets; S3: Based on the two-dimensional training data set, apply the tree regression model for training and prediction, generate and output predicted time series values; and output the feature importance ranking of the tree regression model for decision support.

[0007] In the above method, in step S2, the conversion includes segmenting the time series historical data according to variable window size and step size to generate two-dimensional training data with windows as rows and time steps within the windows as columns.

[0008] In the above method, in step S2, the data of the feature matrix F is preprocessed, and the preprocessing includes data normalization and missing value filling.

[0009] In the above method, in step S3, the tree regression model is a gradient boosting tree model selected from the model group of xgboost, lightgbm or random forest.

[0010] In the above method, in step S3, the training includes a hyperparameter adjustment step, and the hyperparameter adjustment step includes: According to the type of the tree regression model, presetting the hyperparameters to be optimized and their value ranges; Configure sampling rules for grid search or random search based on computing resources and accuracy requirements; K-fold cross validation is used on the training set to evaluate the performance of each hyperparameter combination, with the average index of the validation set as the optimization target; The hyperparameter combination with the best validation index is selected as the final model parameters.

[0011] In the above method, if the sampling rule is grid search GridSearch, the execution of the grid search GridSearch includes the following steps: Discretize continuous hyperparameters by step size to form a multidimensional parameter grid; Train the model for each set of hyperparameter combinations in the parameter grid and record its performance indicators on the validation set; If the indicator is not optimized after N consecutive iterations, the training of the current parameter combination is terminated to reduce the computational overhead; Sort all combinations in descending order by performance index, and select the hyperparameter combination that ranks first for output; If the sampling rule is random search, the execution of the random search includes the following steps: Randomly sample continuous hyperparameters from a uniform distribution or a log-uniform distribution; Set the maximum number of iterations based on computing resources, and randomly select a set of hyperparameters to train the model in each iteration; Increase sampling density for parameter intervals with excellent historical performance; The hyperparameter combination corresponding to the best performance indicator in the iteration is recorded as the final optimization result.

[0012] In the above method, in step S3, the feature importance ranking is calculated based on Gini importance or information gain.

[0013] The above method is applied to scenarios in smart factory projects, including energy consumption forecasting, equipment load forecasting or wastewater treatment cycle forecasting.

[0014] The second aspect of the present invention discloses a time series prediction system based on multi-objective design, comprising a data acquisition module, a data conversion module, a result prediction module and a sorting module; The data collection module is used to collect time series historical data, which includes data points x at multiple time points. t , where t represents the time index; The data conversion module is used to convert the time series historical data into a two-dimensional training data set, which includes a feature matrix F and a multi-target vector T, wherein the dimension of the feature matrix F is m×n, representing m samples and n features, and the dimension of the multi-target vector T is m×k, representing k target columns; wherein, x t is the data point at time t, m is the number of samples, n is the feature dimension, and k is the number of targets; A result prediction module is used to apply the tree regression model to perform training and prediction based on the two-dimensional training data set, and generate and output a predicted time series value; The ranking module is used to output the feature importance ranking of the tree regression model for decision support.

[0015] A third aspect of the present invention discloses an electronic device, comprising: a memory and a processor, wherein the processor is connected to the memory; The memory is used to store programs; The processor calls the program stored in the memory to execute the method provided by the above-mentioned first aspect embodiment and / or any possible embodiment combined with the first aspect embodiment.

[0016] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when run by a computer, executes the method provided by the embodiment of the first aspect and / or any possible embodiment in combination with the embodiment of the first aspect.

[0017] Compared with the existing technology, the present invention has the following advantages: the present invention converts time series historical data into a two-dimensional training data set containing a feature matrix and multi-target vectors, and applies a tree regression model for training, thereby significantly reducing the error propagation risk of multi-step predictions and avoiding the strict dependence of traditional time series models on data stationarity; through the feature importance output natively supported by the tree model, decision-level interpretability is provided while maintaining high computational efficiency, solving the interpretation bottleneck caused by the black box structure of the deep learning model; in the multi-target prediction task, the present invention uses a structured data conversion mechanism to synchronously generate multi-step prediction results, which not only eliminates the error accumulation problem of recursive prediction, but also reduces computational redundancy through parallel output, and ultimately achieves unified optimization of prediction accuracy, computational efficiency and interpretability in resource-constrained scenarios such as smart factory affairs.

[0018] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the method of Example 1 of the present invention.

[0020] Figure 2 This is a flow chart of the method for applying the time series prediction method of the present invention to the ice machine scenario.

[0021] Figure 3 This is a system architecture diagram of Example 2 of the present invention. DETAILED DESCRIPTION

[0022] Example 1 like Figure 1 As shown, a time series prediction method based on multi-objective design includes the following steps: S1: Collecting time series historical data, the time series historical data includes data points x at multiple time points t , where t represents the time index; S2: Convert the time series historical data into a two-dimensional training data set, which includes a feature matrix F and a multi-target vector T, wherein the dimension of the feature matrix F is m×n, representing m samples and n features, and the dimension of the multi-target vector T is m×k, representing k target columns; wherein, x t is the data point at time t, m is the number of samples, n is the feature dimension, and k is the number of targets; S3: Based on the two-dimensional training data set, apply the tree regression model for training and prediction, generate and output predicted time series values; and output the feature importance ranking of the tree regression model for decision support.

[0023] This solution significantly improves forecasting flexibility and interpretability. Compared to existing time series models (such as LSTM), it allows for the selection of a wider range of algorithms (such as tree models), improving model metrics (such as accuracy) in some scenarios (such as energy forecasting). Furthermore, feature importance ranking provides a basis for decision-making, helping engineers optimize plant operations and reduce the risk of misjudgment.

[0024] During implementation, those skilled in the art may refer to the ice machine load prediction scenario. The specific steps are as follows: S1: Collect historical data. For example, in a semiconductor fab, collect load data for a refrigeration machine over the past year (such as hourly recorded temperature and energy consumption), as well as influencing characteristics (such as ambient temperature and equipment operating status). Data point x_t represents the value at time t (e.g., t=1 represents the load at 00:00 on January 1, 2023).

[0025] S2: Convert to a two-dimensional dataset. This isn't a simple rearrangement; rather, it flattens the time series data into a tabular format. For example, assuming the original data is a daily load sequence, a sliding window (e.g., a 7-day window size) is used to construct features: each row of the feature matrix F represents a sample (e.g., the data for day i), and the columns contain features (e.g., the average load or temperature change over the past 7 days). The multi-target vector T includes k target columns (e.g., the predicted load for the next 3 days). This transforms the data from time series into two-dimensional training data, making it easier to train a regression model.

[0026] In terms of training data construction, refer to the attached Figure 2 and Table 1 describes, Figure 2 It shows how to transform a time series (e.g., continuous dates) into a two-dimensional structure (row samples, column features, and targets), which allows one to visualize the data restructuring process.

[0027] An example of training data construction is shown in Table 1 below: Table 1 Example table of training data construction

[0028] S3: Applying tree model training. For example, a LightGBM model is used. After loading a two-dimensional dataset, training is performed: input is a feature matrix F and predictions are made for multiple targets T. Implementation details include model initialization, splitting the training and test sets (e.g., 80% training), and iterative optimization. During prediction, the same transformation is applied to new data (e.g., the current weekly load), outputting future load values ​​and feature importance. After training, the model automatically calculates feature importance (e.g., using information gain) and ranks the outputs (e.g., "ambient temperature" has the greatest impact). Feature importance is the automatically ranked output of the feature importance after successful tree model training. It represents the importance of the feature in influencing the target column. Larger values ​​indicate greater importance. It is often used in conjunction with correlation for attribution analysis; larger importance values ​​indicate greater attribution influence. Engineers adjust ice machine operating parameters accordingly to avoid overloading.

[0029] The entire process is implemented in a standard computing environment (such as Python) without the need for complex coding; tools such as scikit-learn can simplify the implementation.

[0030] In this embodiment, in step S2, the conversion includes segmenting the time series historical data according to variable window size and step size to generate two-dimensional training data with windows as rows and time steps within the windows as columns.

[0031] The conversion uses a sliding window approach with a configurable window size, which improves data adaptability and generalization. Adjustable window sizes (e.g., from 7 to 30 days) accommodate different forecasting needs (e.g., short-term load fluctuations vs. long-term trends) and reduce model overfitting.

[0032] During implementation, skilled practitioners can set window parameters. For example, in the ice machine scenario, the default window size is 7 days; if data volatility is high, this can be adjusted to 14 days. Operation: Define a window function in the data preprocessing script, input time series data, and output a sliding window feature (such as a moving average). This is like "sliding a time frame" to extract relevant data points, and is easily implemented in data processing libraries such as Pandas.

[0033] In this embodiment, in step S2, the data of the feature matrix F is preprocessed, and the preprocessing includes data normalization and missing value filling.

[0034] It's important to note that the feature matrix F in S2 undergoes preprocessing, including normalization and missing value filling, to ensure data consistency and model accuracy. Normalization (such as Min-Max scaling) prevents differences in feature scale from affecting the model, while missing value filling (such as using the mean) avoids data discontinuities and improves prediction reliability.

[0035] Implementation details: After transformation, first check for missing data (e.g., using the isnull() function), fill in missing values ​​with the previous value or mean (e.g., missing load values ​​are filled with the average of the previous week); then normalize to the range 0-1 (e.g., (value - minimum) / range). For example, the ice machine data has a wide temperature range, so normalization makes the model more stable.

[0036] In this embodiment, in step S3, the tree regression model is a gradient boosting tree model selected from the model group of xgboost, lightgbm or random forest.

[0037] During implementation, for example, select LightGBM, initialize model parameters (such as a learning rate of 0.1), and train on a two-dimensional dataset. To predict new data, directly call the model's predict function. Those skilled in the art can quickly deploy this approach without requiring a deep learning framework.

[0038] In this embodiment, in step S3, the training includes a hyperparameter adjustment step, and a grid search or random search method is used to optimize the model parameters.

[0039] During implementation, a parameter grid is defined during the training phase, and GridSearch is used to traverse these combinations and select the optimal parameters. For example, in the ice machine scenario, the adjusted model is more adaptable to seasonal changes. This systematic search for optimal parameters (such as tree depth) avoids the inefficiency of manual parameter tuning.

[0040] Specifically, the hyperparameter adjustment step includes: According to the type of the tree regression model, presetting the hyperparameters to be optimized and their value ranges; Configure the sampling rules of GridSearch or RandomSearch based on computing resources and accuracy requirements; K-fold cross validation is used on the training set to evaluate the performance of each hyperparameter combination, with the average index of the validation set as the optimization target; The hyperparameter combination with the best validation index is selected as the final model parameters.

[0041] For example, if the tree model is LightGBM, the hyperparameter space includes the learning rate (range 0.01–0.2), tree depth (range 3–15), and number of leaf nodes (range 20–100). Grid search needs to traverse all combinations (e.g., learning rate × tree depth × number of leaves, a total of 100 groups), while random search randomly selects 50 groups of combinations for evaluation.

[0042] The execution of the grid search includes: Discretize continuous hyperparameters (such as learning rate) by step size (such as step size 0.05) to form a multidimensional parameter grid; Train the model for each hyperparameter combination in the parameter grid and record its performance indicators (such as root mean square error RMSE) on the validation set; If the indicator is not optimized after N consecutive iterations (e.g., N=5), the training of the current parameter combination is terminated to reduce computational overhead; Sort all combinations in descending order by performance index, and select the hyperparameter combination that ranks first for output.

[0043] For example, define the grid parameter space as {"learning_rate": [0.05, 0.1, 0.15], "max_depth": [5, 8, 10], "num_leaves": [30, 50]}, and generate a total of 3×3×2=18 groups of parameters. Use 3-fold cross-validation for training each combination, and use the average RMSE of the validation set as the evaluation metric.

[0044] The execution of the random search includes: Randomly sample continuous hyperparameters (such as learning rate) from a uniform distribution or log-uniform distribution (for example, the learning rate is sampled from log-uniform(0.01, 0.3)). Set the maximum number of iterations (e.g., 50) based on computing resources, and randomly select a set of hyperparameters for training the model in each iteration; Increase the sampling density for parameter intervals with excellent historical performance (e.g., RMSE below a threshold); The hyperparameter combination corresponding to the best performance indicator in the iteration is recorded as the final optimization result.

[0045] For example, the random search sets the maximum number of iterations to 50, sampling each time from {"learning_rate": loguniform(0.01, 0.3), "max_depth": randint(3, 15)}; when the RMSE of a set of parameters is less than 1.0, 10 additional samples are added to the adjacent interval to improve the local search accuracy.

[0046] The performance index of the validation set is a multi-objective weighted comprehensive score, and the calculation formula is: Score=w1×RMSE+w2×MAE+w3×(1-R²) Among them, w1, w2, and w3 are weight coefficients (default w1=0.5, w2=0.3, and w3=0.2), and RMSE (root mean square error), MAE (mean absolute error), and R² (coefficient of determination) reflect prediction accuracy, robustness, and goodness of fit, respectively.

[0047] Implementation details: In cross-validation, the score is calculated independently for each validation set fold, and the K-fold average is ultimately taken. If the multi-target vector T contains K predicted values ​​(such as the temperature in the next K hours), metrics such as RMSE need to be weighted and summed according to the target dimension.

[0048] Preferably, the hyperparameter adjustment step and model training are performed in a pipeline: The two-dimensional training dataset is split into a training set (70%), a validation set (20%), and a test set (10%) in time series order; Use distributed computing frameworks (such as Spark or Dask) to parallelize the training of multiple hyperparameter combinations. Save a snapshot for each set of candidate models that pass verification, and load the snapshot file to generate predicted time series values ​​when the optimal model is deployed.

[0049] For example, consider 1000 time window data inputs, sort them by timestamp, and use the last 100 windows as the test set. Of the remaining 900 windows, the first 630 are used as the training set, and the last 270 are used as the validation set. The grid search task is distributed to four servers for parallel execution, with each server responsible for training 4–5 sets of parameters.

[0050] In this embodiment, in step S3, the feature importance ranking is calculated based on Gini importance or information gain.

[0051] In implementation, after training, call the model's feature_importances_ attribute to calculate and rank the features (e.g., visualize them using matplotlib). Feature importance ranking is based on Gini importance or information gain calculations, which enhances model transparency and provides a quantifiable basis for decision-making (e.g., identifying key influencing factors).

[0052] In this embodiment, the method is applied to scenarios within smart factory projects, including energy consumption forecasting, equipment load forecasting, or wastewater treatment cycle forecasting. Similar to the ice machine process, these scenarios involve collecting time series data (such as pH values), converting it into two dimensions, and training a model to output cycle forecasts, helping to optimize the process.

[0053] Example 2 like Figure 3 As shown, a time series prediction system based on multi-objective design includes a data acquisition module, a data conversion module, a result prediction module and a sorting module; The data collection module is used to collect time series historical data, which includes data points x at multiple time points. t , where t represents the time index; The data conversion module is used to convert the time series historical data into a two-dimensional training data set, which includes a feature matrix F and a multi-target vector T, wherein the dimension of the feature matrix F is m×n, representing m samples and n features, and the dimension of the multi-target vector T is m×k, representing k target columns; wherein, x t is the data point at time t, m is the number of samples, n is the feature dimension, and k is the number of targets; A result prediction module is used to apply the tree regression model to perform training and prediction based on the two-dimensional training data set, and generate and output a predicted time series value; The ranking module is used to output the feature importance ranking of the tree regression model for decision support.

[0054] The time series prediction system based on multi-objective design provided in this embodiment has the same implementation principle and technical effects as the method embodiment in Example 1. For the sake of brief description, for matters not mentioned in the system embodiment, please refer to the corresponding content in Example 1.

[0055] Example 3 A computer-readable storage medium stores a computer program, which, when executed by a computer, executes the time series prediction method based on multi-objective design described in the above embodiment 1.

[0056] Example 4 An electronic device comprises: a memory and a processor, wherein the processor is connected to the memory; The memory is used to store programs; The processor calls the program stored in the memory to execute the timing prediction method based on multi-objective design as described in Example 1.

[0057] It should be noted that the electronic device may be, but is not limited to, a personal computer (PC), a tablet computer, a mobile internet device (MID), and the like.

[0058] It should be noted that the processor, memory, and other components that may be present in the electronic device are electrically connected to each other directly or indirectly to enable data transmission or interaction. For example, the processor, memory, and other components that may be present may be electrically connected to each other via one or more communication buses or signal lines.

[0059] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referenced to each other.

[0060] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of the systems, methods, and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment, or a portion of code, and the module, program segment, or a portion of code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0061] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0062] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a laptop, a server, a mobile phone, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0063] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A time series prediction method based on multi-objective design, characterized in that: The following steps are involved: S1: Collecting time series historical data, the time series historical data includes data points x at multiple time points t , where t represents the time index; S2: Convert the time series historical data into a two-dimensional training data set, which includes a feature matrix F and a multi-target vector T, wherein the dimension of the feature matrix F is m×n, representing m samples and n features, and the dimension of the multi-target vector T is m×k, representing k target columns; wherein, x t is the data point at time t, m is the number of samples, n is the feature dimension, and k is the number of targets; S3: Based on the two-dimensional training data set, apply the regression model for training and prediction, generate and output the predicted time series value.

2. The method according to claim 1, characterized in that In step S2, the conversion includes segmenting the time series historical data according to variable window size and step size to generate two-dimensional training data with windows as rows and time steps within the windows as columns.

3. The method according to claim 1, characterized in that In step S2, the data of the feature matrix F is preprocessed, and the preprocessing includes data normalization and missing value filling.

4. The method according to claim 1, wherein In step S3, the regression model is a gradient boosting tree model selected from the model group of xgboost, lightgbm or random forest.

5. The method according to claim 4, characterized in that In step S3, the training includes a hyperparameter adjustment step, and the hyperparameter adjustment step includes: According to the type of the tree regression model, presetting the hyperparameters to be optimized and their value ranges; Configure sampling rules for grid search or random search based on computing resources and accuracy requirements; K-fold cross validation is used on the training set to evaluate the performance of each hyperparameter combination, with the average index of the validation set as the optimization target; The hyperparameter combination with the best validation index is selected as the final model parameters.

6. The method according to claim 5, characterized in that If the sampling rule is grid search GridSearch, the execution of the grid search GridSearch includes the following steps: Discretize continuous hyperparameters by step size to form a multidimensional parameter grid; Train the model for each set of hyperparameter combinations in the parameter grid and record its performance indicators on the validation set; If the indicator is not optimized after N consecutive iterations, the training of the current parameter combination is terminated to reduce the computational overhead; Sort all combinations in descending order by performance index, and select the hyperparameter combination that ranks first for output; If the sampling rule is random search, the execution of the random search includes the following steps: Randomly sample continuous hyperparameters from a uniform distribution or a log-uniform distribution; Set the maximum number of iterations based on computing resources, and randomly select a set of hyperparameters to train the model in each iteration; Increase sampling density for parameter intervals with excellent historical performance; The hyperparameter combination corresponding to the best performance indicator in the iteration is recorded as the final optimization result.

7. The method according to claim 1, characterized in that Step S3 also includes outputting a feature importance ranking of the regression model for decision support; in step S3, the feature importance ranking is calculated based on Gini importance or information gain.

8. A time series prediction system based on multi-objective design, characterized in that: It includes data acquisition module, data conversion module, result prediction module and sorting module; The data collection module is used to collect time series historical data, which includes data points x at multiple time points. t , where t represents the time index; The data conversion module is used to convert the time series historical data into a two-dimensional training data set, which includes a feature matrix F and a multi-target vector T, wherein the dimension of the feature matrix F is m×n, representing m samples and n features, and the dimension of the multi-target vector T is m×k, representing k target columns; wherein, x t is the data point at time t, m is the number of samples, n is the feature dimension, and k is the number of targets; A result prediction module is used to apply the tree regression model to perform training and prediction based on the two-dimensional training data set, and generate and output a predicted time series value; The ranking module is used to output the feature importance ranking of the tree regression model for decision support.

9. An electronic device, characterized in that: include: a memory and a processor, wherein the processor is connected to the memory; The memory is used to store programs; The processor calls the program stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a computer, the method according to any one of claims 1 to 7 is executed.

Images