Scenario‐Clustering Dynamic Monotonic Quantile LSTM for Short‐Term Probabilistic Photovoltaic Power Forecasting

ABSTRACT

Short‐term photovoltaic (PV) forecasting supports storage scheduling and renewable‐energy dispatch, but its operational value depends on quantifying uncertainty rather than only estimating an average power trajectory. Weather‐driven PV curves are temporally heterogeneous, and independently trained quantile models often generate crossed or unreliable prediction intervals. To address these challenges, this paper proposes a scenario‐aware monotonic quantile‐learning model, IFCM‐DMCQRLSTM. The IFCM module uses dynamic time warping (DTW) instead of pointwise Euclidean similarity to cluster daily PV profiles by temporal shape, allowing shifted peaks, ramps and weather‐related fluctuations to be grouped into more homogeneous scenarios. DTW‐based similarity is then used to construct scenario‐specific similar‐day samples, on which dedicated DMCQRLSTM models are trained. By embedding composite quantile regression within a long short‐term memory (LSTM) framework and introducing non‐crossing quantile constraints with dynamic quantile adjustment, the proposed model produces monotonic, mathematically consistent multi‐quantile forecasts and adapts interval widths to power fluctuation intensity. Case studies using real‐world data from a 23.4kW PV system in Alice Springs, Australia, show improved point accuracy, reduced quantile crossing and better interval reliability and coverage than benchmark models, confirming its applicability to short‐term probabilistic PV power forecasting.