Under increasingly volatile load conditions, existing short-term load forecasting methods exhibit limited capability in characterizing structural turning features in load sequences, such as peaks and valleys, resulting in degraded prediction accuracy at peak-valley transition moments. To address this issue, a short-term load forecasting method that incorporates peak-valley structural features and a scalar long short-term memory (sLSTM) network is proposed. First, local extrema are extracted from historical load series to construct structured features, including peak-valley state labels, occurrence probabilities, and temporal distances, enabling effective capture of key load transition patterns. Then, an sLSTM is introduced to design a multi-task shared encoding layer, so as to improve the joint modeling capability for load abrupt characteristics and long-term temporal dependencies. Meanwhile, a peak-flat-valley three-classification auxiliary task is constructed to form a multi-task collaborative optimization framework together with the primary load forecasting task. Finally, simulation verification is carried out using publicly available power load data from 2013 to 2014 in a region in China. The results show that the proposed method consistently improves the load forecasting accuracy across different day types while significantly enhancing the prediction effect during peak and valley periods.