Deep LSTM-BiGRU Model for Electricity Load and Price Forecasting in Smart Grids

EasyChair Preprint 8663

Abstract

With the advancement of technology, people are curious to know how much energy they are going to use in the next hour and how much it will cost them. Many accurate prices and load forecasting algorithms are already working but ignore the convergence rate. In the case of STF when an algorithm takes too much time in results formulation, it becomes useless for end users and utilities. We incorporated deep learning techniques as they process a large amount of data quickly and can predict accurate results with a fast computational time. The proposed solution LSTM-BiGRU is formed in combination of LSTM and GRU layers, both are RNN variations and capable of forecasting the best results.  LSTM and GRU are combined in the best possible way to achieve maximum accuracy with a fair computational time. The proposed solution is showing MAPE in load forecasting from 3.12% to 7.42% in different scenarios. Similarly, MAE for price forecasting is calculated between 2.35 to 3.02, and the computational time of the proposed solution in different scenarios is recorded at <1 min. So, a fair tradeoff is maintained between forecasting results and computational time.  In the future, the proposed method can be improved by optimization of proposed hybrid algorithms with evolutionary algorithms, and the use of GPUs and TPU can further decrease the computational time.

Keyphrases: GRU, LSTM, Price forecasting, deep learning, load forecasting, short-term forecasting

@booklet{EasyChair:8663,
  author    = {Aneeza Khalid and Sohail Iqbal and Safdar Abbas},
  title     = {Deep LSTM-BiGRU Model for Electricity Load and Price Forecasting in Smart Grids},
  howpublished = {EasyChair Preprint 8663},
  year      = {EasyChair, 2022}}