The recent success of transformer models for data analysis has resulted in their increased use in varied applications. For some time, transformer models have struggled in Long-Term Time Series Forecasting (LTSF) tasks, where traditional and statistical approaches have been much superior. Recently, a new transformer architecture, iTransformer, has been introduced and has shown state-of-the-art results in various applications. In this paper, we investigate and experiment with the iTransformer model for medium- and long-term forecasting tasks in a real-life application of predicting water temperature in the Murray–Darling Basin, Australia. We compare and evaluate the performance of the iTransformer model to traditional statistical and machine learning approaches. We also investigate different approaches for selecting the historical data for the iTransformer model training and compare the results against a conventional way of selecting the history. Our experimental results show that, in this application, the traditional statistical approach VAR with Seasonal Decomposition (VAR-SD) performs better than the iTransformer model for short-term forecasting range of up to 4 days, but for medium- and long-term forecasting, the iTransformer model performance is superior to both the traditional statistical and machine learning approaches. The results also show that the iTransformer model performs optimally when provided with more recent historical data of approximately the same time length as the forecast horizon.