Modelling Complex Investment Decisions for Renewables with Machine Learning

Abstract

The factors that drive the decision-making process behind private investments in renewables, e.g. solar on roof tops are still somewhat unknown. We aim to develop a more comprehensive model with potential factors from various backgrounds including social, economic and geographic drivers. We use an existing data set of real investments in PV in Germany from 1991 to 2014. These 1.4 million investment decisions are merged with other data sets with information on social, employment, rural/urban characteristics, election results and other potential drivers. The variable of interest is the installed capacity per county. Since the interactions between these variables may be complex, non-linear and are basically not known, we decided to use machine learning statistical methods. In order to increase the robustness of results and to find out which algorithm performs best in terms of model quality, we used Generalized Linear Models (GLM), random forests, gradient boosting and deep neural networks. Model predictions are rather accurate: at the county level the adjusted R2 is 0.65 for GLM, 0.66 for Random Forests, 0.68 for deep neural nets and 0.68 for gradient boosting. Agreement between methods is only decent with deep neural nets calculating a much more balanced model in contrast to gradient boosting. Concerning factor importance for investment decisions, the best two models confirm that the amount of solar insolation received, the absolute number of population per county, and the density and the distinction between urban and rural areas are most relevant