LENARD: Lightweight ENsemble LeARner for MeDium-term Electricity Consumption Prediction

In this work, we propose a lightweight ensemble learner for individual house level electricity consumption prediction. We first implement five different prediction algorithms: ARIMA, Holt-Winters, TESLA, LSTM and Persistence. Among single prediction algorithms, LSTM performs best with 0.0195 MSE value on average. Then, we combine these predictions using neural network based ensemble learner which improves performance of best algorithm (LSTM) on average by 72.84% and by up to 99.13%. Finally, we apply pruning to the weights of our ensemble network to decrease the computational cost of our model. Applying pruning leads to 10.9% less error and 27% fewer number of parameters. We show that our pruned ensemble learner outperforms state-of-the-art ensemble methods.