Probabilistic Load Forecasting with Deep Conformalized Quantile Regression

The establishment of smart grids and the introduction of distributed generation posed new challenges in energy analytics that can be tackled with machine learning algorithms. The latter, are able to handle a combination of weather and consumption data, grid measurements, and their historical records to compute inference and make predictions. An accurate energy load forecasting is essential to assure reliable grid operation and power provision at peak times when power consumption is high. However, most of the existing load forecasting algorithms provide only point estimates or probabilistic forecasting methods that construct prediction intervals without coverage guarantee. Nevertheless, information about uncertainty and prediction intervals is very useful to grid operators to evaluate the reliability of operations in the power network and to enable a risk-based\nstrategy for configuring the grid over a conservative one.\n\nThere are two popular statistical methods used to generate prediction intervals in regression tasks: Quantile regression is a non-parametric probabilistic forecasting technique producing prediction intervals adaptive to local variability within the data by estimating quantile functions directly from the data. However, the actual coverage of the prediction intervals obtained via quantile regression is not guaranteed to satisfy the designed\ncoverage level for finite samples. Conformal prediction is an on-top probabilistic forecasting framework producing symmetric prediction intervals, most often with a fixed length, guaranteed to marginally satisfy the designed coverage level for finite samples.\n\nThis thesis proposes a probabilistic load forecasting method for constructing marginally valid prediction intervals adaptive to local variability and suitable for data characterized by temporal dependencies. The method is applied in conjunction with recurrent neural networks, deep learning architectures for sequential data, which are mostly used to compute point forecasts rather than probabilistic forecasts. Specifically, the use of an ensemble of pinball-loss guided deep neural networks performing quantile regression is used together with conformal prediction to address the individual shortcomings of both techniques.