Dissolved Oxygen (DO) is crucial for sustaining healthy river ecosystems. Recent low DO events in Australia’s Murray-Darling basin have highlighted the need for accurate, long-term DO forecasts. Process-based models are traditionally used but often fail to outperform simple persistence models beyond a few days. Most data-driven DO models have concentrated on high-frequency DO predictions a few days ahead, or on daily averages that may not capture daily DO minima. Building on the success of transformer-based models for high-frequency time series in long-term forecasting, this paper introduces a novel model, ARPatchTST (AutoRegressive Patch Time Series Transformer). It employs a block autoregressive technique, enabling the State-Of-The-Art (SOTA) PatchTST to utilise forecasted block DO concentrations. By leveraging both blocks and patches, ARPatchTST effectively captures intricate local and cyclic patterns from water quality data to improve forecast accuracy of a transformer. Extensive experimental results from three representative monitoring sites in the Murray-Darling basin show that the proposed model outperforms five benchmark models, including the SOTA PatchTST and a persistence model, for 4- or 7-day forecast horizons at 15 min frequency, or daily DO minima. None of the other models can clearly outperform the simple baseline model, persistence, in forecasting daily minima at site 425012, which has experienced several recent fish kill events. Additionally, paired t-tests over 10 independent runs indicate that its improvements over PatchTST are often statistically significant. This work opens several avenues for future research, which are also discussed.