With improved insulation of building envelopes and the use of low-temperature space heating systems, the share of energy use for domestic hot water (DHW) production in buildings has increased significantly, and nearly become the most energy-expensive service in modern buildings. Early prediction of the energy use for DHW is required for many advanced applications such as smart control, demand-side management, and optimal operation of electric or heat storage. However, predicting energy use of the solar-assisted water heating system is more challenging than typical DHW systems, as it is strongly affected by two stochastic phenomena, demand pattern and solar radiation. Given the increasing use of solar-assisted water heating systems, this paper aims to evaluate the potential to predict energy use in such systems using a novel machine learning approach. In this novel model, a Long-Short Term Memory (LSTM) neural network is enhanced by (1) implementing the attention mechanism, a recent development in deep learning inspired by human vision to pay selective attention to the input data, and (2) decomposition of input data into sub-layers. The performance of simple LSTM neural network, Attention-based LSTM neural network (ALSTM) and Attention-based LSTM using decomposed data (ALSTM-D) are compared to a Feed-Forward neural network as a baseline model. Results show that LSTM, ALSTM and ALSTM-D models have a Mean Absolute Error (MAE) of 25%, 28% and 41% lower than Feed-Forward model, respectively. These results indicate the superior performance of the proposed ALSTM-D model over conventional models for solar-assisted DHW systems.