Nearest-Neighbor Based Non-Parametric Probabilistic Forecasting with Applications in Photovoltaic Systems
For practitioners in renewable energy forecasting, this offers a simpler way to create probabilistic forecasts without specialized loss functions, but the results are only 'acceptable' and not state-of-the-art.
This paper proposes a simple method for interval forecasting by combining quantile regressions built via k-nearest-neighbors-based data transformation, avoiding the pinball loss. Applied to photovoltaic power forecasting, the method yields acceptable interval forecasts using polynomial models, neural networks, and support vector regressions.
The present contribution offers a simple methodology for the obtainment of data-driven interval forecasting models by combining pairs of quantile regressions. Those regressions are created without the usage of the non-differentiable pinball-loss function, but through a k-nearest-neighbors based training set transformation and traditional regression approaches. By leaving the underlying training algorithms of the data mining techniques unchanged, the presented approach simplifies the creation of quantile regressions with more complex techniques (e.g. artificial neural networks). The quality of the presented methodology is tested on the usecase of photovoltaic power forecasting, for which quantile regressions using polynomial models as well as artificial neural networks and support vector regressions are created. From the resulting evaluation values it can be concluded that acceptable interval forecasting models are created.