Comparative analysis of the outcomes of differing time series forecasting strategies

Abstract

Time series forecasting of data from various domains has become an increasingly interesting research subject in recent times. Prediction of future sample values is the main goal of time series forecasting. There are mainly two classes of time series forecasting, namely, single step and multi-step forecasting. There are various machine learning approaches that can be used along with various forecasting strategies which exist, some of which can be utilized in and by time series forecasting. However, inappropriate use of strategies can limit the applicability of the techniques to real world problems. This paper provides a review of time series forecasting, a classification of time series forecasting and also approaches and strategies of time series forecasting. It also demonstrates how inappropriate use of point to point rolling forecast strategy for forecasting could lead to unrealistic outcomes and how a multiple step forecasting strategy called Direct H step strategy could help to overcome this issue. Comparative analysis of two strategies using Auto-Regressive Integrated Moving Average (ARIMA) approach are demonstrated.

Multi-step Time Series Forecasting of Electric Load Using Machine Learning Models

“…The Dickey-Fuller (ADF) test [4] [32] was applied and it was found that the time series data of all three datasets is non-stationary whereas it is important to have the time series data to be stationary for application of the ARIMA forecasting model [4] [32]. Seasonal differencing was applied on the three datasets to make the time series stationary as outlined in [4] [32]. It was found that before doing the first order differencing the ADF Statistic was found to be more than the 1% critical value and the p-value was close to 0.05 so the null hypothesis of the ADF test could not be rejected.…”

“…Whereas after the seasonal difference ADF Statistic to be found it was less than the 1% critical value and the pvalue was much lower than 0.05 which suggests that the d parameter of the ARIMA model should at least be a value of 1 for better performance. Next, for the other two parameters p and q of the ARIMA model, the Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) were computed and plotted following the procedure as outlined in [4] [32]. Examining the ACF and PACF plots, the parameters p and q were identified for all three datasets.…”

“…Duration of electric load forecasting can be mainly classified into two categories, short-term load forecasting and long-term load forecasting [3]. Single-step forecasting and multi-step forecasting are useful for short-term and long-term load forecasting respectively [4]. Several machine learning approaches have been considered for forecasting which can be divided into two broad categories, statistical techniques and soft computing techniques [5].…”

Multi-Step Short-Term Power Consumption Forecasting Using Multi-Channel LSTM With Time Location Considering Customer Behavior

“…In general, power consumption forecasting contains three categories: short-term forecasting (STF), medium-term forecasting (MTF), and long-term forecasting (LTF) according to the forecasting duration [3]. Moreover, it could be classified into two categories according to forecasting steps: one-step forecasting and multi-step forecasting [2], [4]. One-step forecasting employs historical power consumption related variables to predict the next one-step.…”

“…i.e., B. Wolff et al [7] used support vector regression (SVR) for photovoltaic (PV) power forecasting with numerical measurement weather and cloud motion data. One of the most popular model-autoregressive integrated moving average (ARIMA) was employed for power forecasting in [4], [8], [9]. The SVR regression-based method aims to find one regression plan and enable all the data of a set to the nearest distance to the plane.…”