PP-TS time series prediction model based on heuristic search and ensemble method to make prediction more accurate

Time-series data is ubiquitous in various industries and domains, such as measurements from IoT sensors, hourly sales performance, stock prices in the financial sector, etc. are examples of time-series data. Time series forecasting is the use of historical multidimensional data for statistical analysis to infer the future development trend of things.

In order to speed up the process of enterprise intelligent transformation and reduce the threshold of timing technology application, Flying Paddle continues to polish product technology, and launched a high-precision timing model PP-TS based on heuristic search and integrated learning, which has been verified on the power scene data set . Increase over 20%.

PP-TS officially launched the Paddle AI kit PaddleX today! The source code is all open! You can explore it on the cloud of AI Studio (Galaxy Community) or on the local side of PaddleX! Try to combine it with real business scenarios! In the toolbox mode, you only need to provide historical data in a scenario, and PP-TS can accurately predict the data situation in the future for you in that scenario.


pictureassessment test

In addition to PP-TS, Paddle also provides 8 industry-leading time series prediction methods, namely TimesNet, TiDE, PatchTST, DLinear, RLinear, NLinear, Nonstationary Transformer and XGBoost for your comparison and use.

PP-TS core idea

Based on the integrated method, PP-TS forecasting is more accurate than traditional time series forecasting With the advent of the 5G era, enterprises are gradually entering a new stage of digital transformation, facing more and more complex time series forecasting scenarios, such as equipment remaining life forecast, power load forecasting, etc. In complex time series forecasting scenarios, characteristics such as long time series, multi-variables, and non-stationarity seriously affect the accuracy of model predictions, and put forward higher requirements for time series forecasting tasks. Therefore, we adopt an integrated method to achieve better forecasting performance by selecting and merging multiple single-prediction models. In the power scenario, the accuracy of PP-TS is better than that of traditional models, and the prediction error is reduced by an average of 30%
Through heuristic search, the number of searches is reduced

Different single models have different capabilities. For example, Non-Stationary Transformer is optimized for non-stationary data scenarios, and TimesNet has stronger expressive capabilities through multi-period decomposition. It can be seen that the integration of different algorithm combinations will produce different prediction performance, so what model to choose can achieve the best effect, and how to improve the selection efficiency in the process of selecting model combination, these are the problems that need to be focused on and solved. Therefore, we propose PP-TS, which can reduce the number of combination of selected models through heuristic search and selection model integration. There are 2^8 kinds of integration combinations. Through heuristic search, the number of searches can be reduced to about 30 times, while ensuring that the integrated model can achieve the best accuracy.

Interpretation of PP-TS key technical points

The overall technical framework of PP-TS with stronger capabilities is shown in the figure below:


PP-TS mainly conducts in-depth exploration from three perspectives, mainly including:

  • Basic single model: deep models generally have strong fitting ability, Transformer-based methods are good at capturing long-term dependencies, and machine learning methods have better interpretability. PP-TS chooses the combination of cutting-edge deep models and traditional methods, including TimesNet, TiDE, PatchTST, DLinear, RLinear, NLinear, Nonstationary Transformer, and XGBoost.
  • Heuristic search: Model whether a single model is selected as a 0/1 problem, evaluate the accuracy of the selected combination through the genetic algorithm, and screen the optimal combination by selecting cross-mutation evolution.
  • Model integration: the selected models are integrated, the results are fused, and the method with the best accuracy is obtained.

How to customize personalized PP-TS

In real business, it is generally not recommended to use the general version of PP-TS directly, but needs to be specifically optimized and adapted to the data type in the business scenario to achieve high enough accuracy and stability to meet real business needs. So, how to create a personalized PP-TS? Let's take a look at them one by one.
Create the PP-TS model production line
The flying paddle AI kit PaddleX has been launched in AI Studio (Galaxy Community), the current entrance is in the model library , you can find PP-TS here, read its introduction document, and finally create your own PP -TS model production line. The link to the AI ​​Studio (Galaxy Community) model library is as follows: https://aistudio.baidu.com/aistudio/modelsoverview?supportPaddlex=true&sortBy=weight!

Prepare evaluation data and effect verification

In order to objectively evaluate the effect of PP-TS, it is recommended that you prepare a certain amount of evaluation data from business scenarios for quantitative comprehensive evaluation. For example, industry data for the past three months can be prepared, and fields of interest and their true values ​​can be marked according to business requirements .

After evaluating the general PP-TS effect, the result can be used as a baseline to guide subsequent optimization work for business scenarios.

One more thing , in the future, in the Paddle AI suite PaddleX, everyone can not only develop their own models, but also make joint contributions and share revenue with the platform!

The Lianchuang model can not only realize technology, but also give individual developers a sense of accomplishment, and attract traffic and attention for enterprise developers. It can be said that good things come in pairs! Moreover, with a large number of users, valuable feedback can be collected, and contributors can be encouraged to further optimize the model, thereby attracting more users, which can be described as a double helix~

In order to protect the intellectual property rights of contributors, we will provide a complete encryption and authentication mechanism. Contributors only need to access the encryption and authentication capabilities according to our documentation, and you can contribute models with confidence! For more details about Lianchuang, please pay attention to the follow-up update of the Paddle AI kit PaddleX!

PP-TS in PaddleX Paddle AI Kit


PP-TS GitHub


Guess you like

Origin blog.csdn.net/PaddlePaddle/article/details/132407872