With the increase in volatile power generation from photovoltaic systems, accurate forecasting is becoming essential to mitigate undesired effects on the electrical grid. However, the accuracy of forecasts is significantly impaired by uncertainties in weather predictions and long-term recurring effects on photovoltaic yield, such as seasonal patterns or frequent shading. This paper proposes a clustering-based feature engineering method to extract features that capture otherwise undetectable trends. The resulting descriptive feature is then integrated into a post-processing model to improve forecast accuracy. By incorporating this engineered feature, the error of photovoltaic power forecasts can be reduced. Specifically, the normalized mean absolute error decreases from 8.76% for the initial forecast to 7.67% after applying the proposed post-processing approach, demonstrating the effectiveness of the method.