A comparative study of LSTM and ARIMA for energy load prediction with enhanced data preprocessing,

Innocent Mpawenimana (1) , Alain Pegatoquet (1) , V. Roy (2) , Laurent Rodriguez (1) , Cécile Belleudy (1)

1 LEAT - Laboratoire d'Electronique, Antennes et Télécommunications
2 Mines Paris - PSL (École nationale supérieure des mines de Paris)

Energy load prediction plays a central role in the decision-making process of energy production and consumption for smart homes with systems based on energy harvesting.However, forecasting energy load turned out to be a difficult problem since time series data used for the prediction involve both linear and non-linear properties. In this paper, we proposed a system which can predict a daily future energy load in a smart home based on LSTM and ARIMA models. To improve the energy load forecasting accuracy, we propose a new data preprocessing algorithm called STDAN (Same Time a Day Ago or Next) to fill the missing values. This technique is compared with well-known techniques using previous or mean values. A comparison between LSTM and ARIMA is provided for short and medium-term load forecasting. Results show that LSTM outperforms ARIMA in all cases. Finally, we also evaluated our training model based on LSTM with a new data set and the model provides an around 80% accuracy.

Mots clés: ARIMA; Forecasting; Data Preprocessing; Electric Load; Deep Learning; Machine Learning; LSTM

Domaines: Systèmes embarqués; Architectures Matérielles [cs.AR]; Performance et fiabilité [cs.PF]

Innocent Mpawenimana, Alain Pegatoquet, V. Roy, Laurent Rodriguez, Cécile Belleudy. A comparative study of LSTM and ARIMA for energy load prediction with enhanced data preprocessing,. 15th IEEE Sensors Applications Symposium (SAS), Mar 2020, Kuala Lumpur, Malaysia. pp.1-6, ⟨10.1109/SAS48726.2020.9220021⟩. ⟨hal-02456709⟩