Time-Series Forecasting of Seasonal Data Using Machine Learning Methods

Abstract

The models for forecasting time series with seasonal variability can be used to build automatic real-time control systems. For example, predicting the water flowing in a wastewater treatment plant can be used to calculate the optimal electricity consumption. The article describes a performance analysis of various machine learning methods (SARIMA, Holt-Winters Exponential Smoothing, ETS, Facebook Prophet, XGBoost, and Long Short-Term Memory) and data-preprocessing algorithms implemented in Python. The general methodology of model building and the requirements of the input data sets are described. All models use actual data from sensors of the monitoring system. The novelty of this work is in an approach that allows using limited history data sets to obtain predictions with reasonable accuracy. The implemented algorithms made it possible to achieve an R-Squared accuracy of more than 0.95. The forecasting calculation time is minimized, which can be used to run the algorithm in real-time control and embedded systems.

Forecasting of International Flights Passenger at Soekarno-Hatta Airport using The Triple Exponential Smoothing Method

“…If the historical identification of the actual demand data shows seasonal fluctuations, it is necessary to adjust for seasonal effects by calculating the seasonal index. As an example to explain the effect of seasonality using seasonal index numbers (Himawan & Silitonga, 2020;Jonnius, 2017;Kramar & Alchakov, 2023;Triana, 2015).…”

“…To determine the smoothing constant α if the data used has an unstable historical pattern then the value of the smoothing constant α has a value close to 1, but if the data used has a historical pattern that does not fluctuate or is relatively stable then the value of the smoothing constant α is close to 0 Forecasting using the triple exponential smoothing method is to calculate exponential smoothing, trend smoothing, and seasonal smoothing. After the three factors have found their smoothing values, the last step is to forecast the data for the upcoming p period using the formula: 𝑌 ̂𝑡+𝑝 = 𝑆 𝑡 + 𝑝𝑏 𝑡 + 𝑙 𝑡−𝐿+𝑝 (15) With 𝑌 ̂𝑡 = the value to be forecast and p = the time period to be forecast This method uses three smoothing constants namely , 𝛽, and 𝛾 with a weighting value ranging from 0 to 1 (Kramar & Alchakov, 2023;Santoso & Kusumajaya, 2019;Suryani et al, 2023).…”

“…by N is the amount of data observed, 𝑌 ̂𝑡 is forecast of t-th, and Yt is t-th data. A data model will have very good performance if the MAPE value is below 10% (Kramar and Al, 2023).…”

Seasonal Data Cleaning for Sales with Chase Demand Strategy

“…Ref. [10] focuses on the forecasting of seasonal time series data using machine learning methods. They discuss the significance of accurately forecasting parameters with seasonal variability and the application of various machine learning techniques to address seasonal data.…”

Optimized Website Traffic Forecasting with Automatic Models and Optuna

“…In contrast, linear forecasting models are simpler and assume that the trend will continue in a straight line. Companies need to analyze their website traffic patterns and use the most appropriate forecasting models to predict traffic levels accurately [4]. This information can be used to allocate computer resources effectively and forecast future income and advertising growth.…”