Abstract: DESCRIPTION: Forecasting the long term values of a time series data is crucial for planning. How do you make use of a Recurrent Neural Network when you want to compute an accurate long term forecast? How can you capture short and long term seasonality? Can you learn small patterns from the data that generate the big picture? This session will provide a scalable technique addressing these questions.
ABSTRACT:Neural networks are powerful function approximators that can capture and represent linear and non-linear variations in the data. Recurrent Neural Networks (RNN) extend this approach by taking into account the sequential nature of the data. This sequential relationship is captured by creating an internal description of the input, which is computed by not only using the current observation in a series but also the current representation of all the points that have been observed so far.
In this talk I will present an approach for training an RNN on time series data that exhibit short and long term seasonality to generate multiple time steps ahead forecast. In particular I will present how to transform a long time series input into shorter sequences that retain the patterns relevant for capturing the multiple modes of seasonality and the variability. This approach allows one to multi-thread and distribute the training process.
I will demonstrate the long term forecast performance using SAS’ solar farm production data and show ways to incorporate derived features to enhance the time series for better accuracy. I will present results that generate much better long term forecast accuracy than the traditional time series forecasting methods. This technique is extensible to practical applications like forecasting consumption demand, predicting stock markets and weather conditions.
Bio: Mustafa Kabul is a data scientist in the Artificial Intelligence and Machine Learning R&D at SAS, where he leads innovative projects for SAS’s next-generation AI-enabled analytics products, including applications of deep learning. His current focus is on applying deep reinforcement learning to operational problems in the CRM and IoT spaces. An operations research expert working at the interface of machine learning and optimization, previously, he developed distributed, large-scale integer optimization algorithms for marketing optimization problems. Ever the optimization enthusiast, Mustafa always looks into ways to improve the algorithms. Nowadays his favorites are the distributed stochastic gradient and online learning methods. Mustafa holds a PhD from the University of North Carolina at Chapel Hill, where his research focused on game theory models of supply chains selling to strategic customers.