Heterogeneity in the Adoption of Photovoltaic Systems in Flanders

Abstract

We study the determinants of PV adoption in the region of Flanders (Belgium), where PV adoption reached high levels during 2006-2012, because of active government intervention. Based on a unique dataset at a very detailed spatial level, we estimate a Poisson model to explain the heterogeneity in adoption rates. We obtain the following findings. First, local policies have a robust and significant impact on PV adoption, providing indirect evidence that the larger regional incentives formed the basis for the strong development of PV adoption in the region. Second, there is a strong unconditional income effect, implying a Matthew effect in the subsidization of PVs. Our third finding is however that this income effect is largely driven by the fact that wealthier households are more likely to adopt because they tend to be larger (and hence higher users), are more frequent house owners (who capture more of the benefits), or own houses that are better suited for PV. We can thus identify the channels through which wealthier households are more likely to benefit from the PV support. Finally, we identify the importance of several housing characteristics: PV adoption tends to be more likely in larger and in more recently built houses. In several extensions, we consider the determinants of the average size of installed PVs, and the differential impact of certain variables over time.

PV adoption: the role of distribution tariffs under net metering

“…For this purpose, we consider the approach used in the literature by Bollinger and Gillingham (2012), Muller and Rode (2013) or Graziano and Gillingham (2014). In regression (12), we control in addition for the accumulated number of previously installed PV's in the municipality, in addition to the estimation strategy used in our benchmark regression (1). We observe that this does not have an impact on our main result.…”

“…Statistical significance: * p < 0.1, * * p < 0.05, * * * p < 0.01 Table 5 presents various robustness checks with respect to different estimation strategies. In regression (12) and (13), we first control for the presence of social interaction peer effects in the diffusion of PV panels. These social drivers help overcome non-monetary barriers to adoption via localized knowledge sharing among neighbors.…”

“…Hence, we believe that reverse causality is not a concern in our particular institutional setting. 12…”

DeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States

“…Models that estimate deployments from socioeconomic and environmental variables are key for decision making by regulatory agencies, solar installers, and utilities. Studies have focused on either utilizing surveys [17][18][19][20][21][22][23] or data-driven approaches [14][15][16][24][25][26][27][28] at spatial scales ranging from county-to state-level models, achieving in-sample R 2 values between 0.04 and 0.71. The models are typically linear 28 or log-linear 27 and utilize less than 10,000 samples for regression.…”

“…These 94 factors are the largest set of factors we can collect for all census tracts and part of them have been also utilized and reported in previous works. [14][15][16][24][25][26][27][28] Among all predictive models, the Random Forest-based model, called SolarForest, achieves the tier-1 out-of-sample R 2 value of 0.72 in the 10-fold cross-validation, which is even higher than the in-sample R 2 values of any other models in previous works. [14][15][16][24][25][26]28 SolarForest is a novel machine learning-based hierarchical predictive model that postulates census tract level solar deployment as a two-stage process: whether tracts contain solar panels or not, and, if they do contain them, the number of systems per household is decided ( Figure 6A).…”