To address the deviation between medium- and long-term distribution network planning and actual conditions caused by high penetration of renewable energy and the uncertainty of source-load output, a planning-state source-load joint scenario generation method considering temporal feature decomposition is proposed. First, the seasonal and trend decomposition using LOESS (STL) method is used to decompose photovoltaic (PV)-load time series data into periodic, trend, and residual components, each with distinct characteristics. Next, the TimeMixer deep learning model is employed to forecast the seasonal and trend components for the planning year. This model captures temporal information across multiple scales and effectively integrates micro-periodic patterns with macro-trend information. Meanwhile, for the residual components, which contain random and unpredictable features, a time-varying Copula dependence modeling method is adopted to characterize the historical PV-load correlation. By combining the forecasted components, PV-load joint scenarios are generated, and various clustering methods are applied to representative scenarios for analysis. Finally, a case study using integrated PV-load data from Belgium’s Elia operator verifies that the proposed method can generate planning scenarios that reflect future growth and significantly improve scenario accuracy.