1 A Hybrid Deep Stacked LSTM and GRU for Water Price Prediction Abdullahi Uwaisu Muhammad 1.∗ , Adamu Sani Yahaya 2,∗ , Suhail Muhammad Kamal 2 , Jibril Muhammad Adam 1 , Wada Idris Muhammad 3 , Abubakar Elsafi 4 1 Department of Computer Science, Federal University Dutse, Nigeria 2 Department of Information Technology, Bayero University, Kano, Nigeria. 3 Department of Soil Science and Land Resource Management, Federal University, Wukari, Nigeria 4 Software Engineering Department, College of Computer Science and Engineering, University of Jeddah, Jeddah, Saudi Arabia ∗ Corresponding author: asyahaya.it@buk.edu.ng; uwaisabdullahi87@yahoo.com Abstract—Water pricing and freshwater scarcity is an emerg- ing global issue, a topic of debate among researchers, households and water utility managers. This is due to the fact that, the process can provide early warning signs as well as assisting water utility managers to make proper decisions on control and management of the scarce water resources through implementing water pricing policies, ensuring proper water allocation, water- use restriction as well as water production. In this paper, we presented a two-step methodology coupled stacked LSTM+GRU models while analyzing their relative performance to our ref- erence models i.e. stacked LSTM and GRU for long term water price Prediction. It is thought that, the coupled Stacked LSTM and GRU models to exploit building of higher level of representation of the input sequence data while creating a higher level of abstraction on the final results. The GRU on the other hand assists in solving the vanishing gradient problems. The experimental results obtained from this research work indicates our coupled (Stacked LSTM+GRU) with supervised learning to significantly outperform our reference models for water price Prediction. Index Terms—Water Price Prediction, Stacked LSTM+GRU, Stacked LSTM, GRU, Accuracy. I. I NTRODUCTION A very important role in the planning, management of urban water demand is efficient and reliable water pricing. Water pricing helps water utility managers to make proper decision on the operation and management of existing water supply through effective and efficient water pricing policies, proper allocation of scarce water re-sources, effective water use restrictions and water treatment and recycling. Water withdrawals from lakes, streams, rivers, ponds and dams are at increasing rate due to rapid population growth, industrializa- tion, improved standard of living, consumer behavior pattern and migrations [1], [2]. As such accurate water price prediction could be of great economic use and might have positive effects to lives and properties of future generations. This process also requires a better understanding of the influence of popula- tion, climate change and consumer behaviour pattern through awareness. Over the years, the number of individuals, NGOs and governmental bodies investing in the proper treatment, discharge and marketing of fresh water have increased. This is due to the tempting nature, thus worthy the exploring of the variables likely affecting water price forecasting in the future. An inexperienced investor who might just check the historical fluctuations of water stock prices might not be really aware of the risks involved, as they could be misleading because they do not normally provide the details of the factors influencing the rise or fall of water prices at the market. The aim of this research is to assess the prediction performance of Stacked LSTM + GRU for daily water price forecasting using Aqua water company which are specialized in the production, purchase, distribution and sales of water resources in United States. [3] The performance of GRU network was compared with conditional ANN and seasonal ARIMA models for short term water demand forecast. The results show that deep learning method improves the performance of water demand prediction while correlation module enhances the performance of ANN and GRU models. They suggest future research to consider using other techniques e.g. auto-encoder and convolution transformation to further improve the models performance as well as the use of real-time monitoring data. [4] investigation of trading models for historical data, show a consistent profit growth by using Delphi and the calculation of the compatibility of forecasting of LSTM-based recurrent neural network. The Delphi method improves boundaries of predictions. They suggest future investigations in a longer trading period with a different input data of exchange rates. [5] developed a high performance information extraction frame- work for e-cigarette social media safety surveillance using deep neural network method. The Bi-LSTM RNN model achieved much higher recall, resulting in higher F-measure than strong statistical learning lexicon baseline. They suggest future research to focus on finding the best configuration of model parameters. To overcome the problem of global water scarcity owing to rapid population growth, industrialization, urbanization as well as over extraction of water resources. We explored the literature, performed the practical and compared the relative performance of hybrid stacked LSTM and GRU to our benchmark models GRU, and stacked LSTM models using daily stock water price data set. The overall performance of the models in terms of mean absolute error (MAE), mean 978-1-7281-5467-1/20/$31.00 c 2020 IEEE Authorized licensed use limited to: Hohai University Library. Downloaded on November 29,2020 at 10:47:20 UTC from IEEE Xplore. Restrictions apply.