New approach to evaluating predictive models of photovoltaic systems

Abstract Performance models are used for forecasting the energy output of photovoltaic systems. The ability of a model to accurately predict the output must be carefully assessed because the capital cost for photovoltaic system implementation is tied to the reliability of the forecast. The customary approach to model evaluation is based on the root-mean-square-error, mean-absolute-error and mean-bias-error. The values of these statistics strongly depend on the scale of the quantities they compare and, therefore, can be misleading. This paper presents a multifaceted approach to model evaluation, comprising qualitative and quantitative assessment, thus increasing the reliability of the evaluation outcome. The qualitative assessment is based on graphical residual analysis. The quantitative assessment is based on two new, scale independent, metrics. The results show good agreement between the outcomes of the qualitative and the quantitative assessment, when it is done with the new metrics, but a discrepancy between the two, when the traditional metrics are used. The new metrics can be used to assess the forecasting accuracy of different performance models of the same photovoltaic system as well as to compare the accuracy of models of different systems. The new approach is applied in the evaluation of models of two photovoltaic systems and the results are compared with the results that would have been obtained, had the customary metrics been used.