Spatio-temporal modelling of diffusion of electric vehicles and solar photovoltaic panels: an integrated agent-based and artificial neural networks model.

Abstract

This work investigates the spatio-temporal patterns of EV and PV adoption, attending to the need of informing the management of the distribution network with spatially explicit estimations of the EV and PV adoption rates. The research assesses the strengths and weaknesses of current modelling approaches and finds three recurrent techniques: agent-based model (ABM), the spatial regression (SR) and the Poisson model. The research addresses the limitations of the current modelling approaches to characterise the different factors that drive the decision-making for the adoption of these technologies: spatial and temporal dependence and social dynamics. The framework addresses some of the limitations of ABMs that use a rule or equation-based decision-making, by adopting an aggregated agents’ definition and using artificial neural networks (ANN) as decision-making criteria. The integrated model provides a more realistic characterisation of the decision-making and its evolution over time, moreover, the results can inform network operators and policymakers explicitly about the location and pace of EV and PV adoption. The research develops a spatio-temporally explicit ABM that accounts for: (i) spatial and (ii) temporal dependence, (iii) peer-effect, (iv) spillover effect and (iv) preferences towards other technologies. The development of the model follows a sequential approach to managing the complexity of constructing this new approach. First, two autoregressive models are developed to analyse the adoption patterns of EV and PV patterns, using the postcode and monthly data resolution. The temporal validation uses the Mean Absolute Percentage Error to measure the model’s capability to replicate the time-series of the adoption rates. The spatial validation compares the actual and estimated spatial pattern of adoption by calculating the Moran’s I index. Besides, the results are benchmarked against the Bass model, a commonly used tool for this purpose by ABM experts. The results show that in most of the cases the ABM and ANN integrated models perform better than the Bass model especially for those months with high fluctuations in the adoption rates. These models can estimate upmost three months with an accuracy higher than 80%, however, the models present a significant accumulation of errors that limits the results for a longer forecast. To reduce the error accumulation and produce a longer forecast, the autoregressive PV model is extended by including socioeconomic variables. The resulting model improves the performance by 5% by the incorporation of variables including income, electricity consumption and average household size. Lastly, the framework combines the EV and PV autoregressive models with a view to characterising the exchange of knowledge between EVs and PVs. This reflects the influences of owning one of those technologies on the preference for the second technology. The exchange of knowledge improves the performance of the model significantly with results above 80% of accuracy for eight months into the future. Given the high spatial resolution of the model, the results may help to design policies that recognise the socioeconomic differences within a geographical area. The research shows how the results can inform the management of the distribution network, by considering the worst-case scenario where the PV generation surplus is injected to the grid, and where the entire fleet of EVs are charged at home during the night. Also, the results of the hot spot analysis can inform the network operators about the emergence of clusters of EVs and PVs in the future. The research finds that a spatio-temporally explicit ABM can characterise the EV and PV adoption process at the aggregated level, which also accounts for social effects. Such a model can also integrate heterogeneity amongst the population, whilst being resilient to changes in the size of the study area. The research also produces data-driven insights into the spatio-temporal patterns of EV and PV adoption, and how the adaptive capabilities of the ANN address some of the limitations of the ruled-based ABM. Lastly, the research finds that knowledge exchange takes place between the EV and PV adoption process. These findings are relevant to other low carbon technologies and for the modelling of other sociotechnical systems.