Predictive analytics beyond time series: Predicting series of events extracted from time series data

Realizing carbon neutral energy generation creates the challenge of accurately predicting time-series generation data for long-term capacity planning and for short-term operational decisions. The key challenges for adopting data-driven decision-making, specifically predictive analytics, can be attributed to data volume and velocity. Data volume poses challenges for data storage and retrieval. Data velocity poses challenges for processing the data near real time for operational decisions or for capacity building. This manuscript proposes a novel prediction method to tackle the above two challenges by using an event-based prediction in place of traditional time series prediction methods. The central concept is to extract meaningful information, denoted by events, from time-series data and use these events for predictive analysis. These extracted events retain the information required for predictive analytics while significantly reducing the volume of the velocity of data; consequently, a series of events present the information at a glance, effectively enabling data-driven decision-making. This method is applied to a data set consisting of six years of historical wind power capacity factor and temperature measurements. Deploying five deep learning models, a comparison is drawn between classical time-series predictions and series of events predictions based on computational time and several error metrics. The computational analysis results are presented in graphical format and a comparative discussion is drawn on the prediction results. The results indicate that the proposed method obtains the same or better prediction accuracy while significantly reducing computational time and data volume.

Automated summary

This study proposes an event-based prediction method for data-driven decision-making, which effectively reduces computational time and data volume while maintaining high prediction accuracy. The method is applied to a wind power dataset, and the experimental results show its superiority over traditional time series prediction methods.