Short-term　water　demand　forecasting　is　essential　for　ensuring　the　sustainable　use　of　water　resources.　The　accuracy　of　water　demand　forecasting　directly　impacts　the　rationality　of　water　resources　management　and　the　effectiveness　of　scheduling.　Therefore,　it　is　vital　to　accurately　forecast　water　demand　across　various　timescales.　Based　on　this　motivation,　we　propose　an　improved　patch　time　series　Transformer　(PatchTST)　model　to　forecast　the　multi-timescale　short-term　water　demand.　By　introducing　relative　positional　encoding　(RPE),　the　model　effectively　learns　the　relationships　between　tokens.　The　model　combines　the　global　token　information　capture　ability　of　the　self-attention　mechanism　with　the　local　token　information　capture　ability　of　the　convolutional　network　to　enhance　feature　extraction　abilities.　Additionally,　the　model　integrates　the　advantages　of　patch-wise　and　series-wise　representation,　enabling　it　to　simultaneously　capture　both　local　and　global　dependencies　in　time　series.　We　utilize　historical　data　collected　from　district　metering　area　to　experimentally　validate　the　effectiveness　of　the　proposed　model.　Compared　with　one-dimensional　convolutional　neural　network　(1D-CNN),　long　shortterm　memory　(LSTM),　Transformer,　DLinear,　and　PatchTST　models,　our　model　demonstrates　superior　performance　across　all　five　forecasting　scales.　Finally,　the　effectiveness　of　the　proposed　design　is　further　validated　through　ablation　experiments.