While forecasting key performance values of EPES assets has been commonly tackled over time, effective predictive maintenance approaches will require the detection of of faulty operating conditions especially when these faulty samples where not present in the training set.  Achieving this requires a perfect balance of not overfitting deep neural network models when there is abundance of data  as well as addressing underfitting the model when there is limited data—both with their consequent effects in forecasting timeseries data. The solution stacks up the predictions of independent models leveraging attention concepts while selecting the best features that were best predictors of the PV performance.