Accurate forecasting of photovoltaic (PV) power output plays a critical role in the reliable integration of solar energy into modern power grids and in the optimal operation and management of renewable energy systems. Nevertheless, achieving high prediction accuracy remains challenging due to the inherent variability and stochastic nature of meteorological conditions, including solar irradiance, ambient temperature, and atmospheric dynamics. These uncertainties significantly affect PV power generation and necessitate the use of advanced data-driven modeling techniques capable of capturing complex nonlinear and temporal relationships. In this study, a comprehensive benchmarking analysis of four widely used deep learning architectures, Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), is conducted for photovoltaic power forecasting. The evaluated models represent both feedforward and recurrent learning paradigms, allowing for a systematic comparison of their predictive capabilities in modeling PV power dynamics. The models are trained and tested using real-world operational data collected from a grid-connected PV plant in China, comprising meteorological variables and historical PV power output measurements. Prior to model training, appropriate data preprocessing and normalization steps are applied to ensure consistency and robustness. PV power output is predicted at time t using a supervised learning framework, enabling short-term forecasting under real operating conditions. Furthermore, the proposed framework is inherently flexible and can be extended to medium-term forecasting horizons of up to 15 days by integrating Numerical Weather Prediction (NWP) data as external inputs. Model performance is quantitatively assessed using widely accepted evaluation metrics, including Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²), ensuring a fair and transparent comparison across all architectures. The experimental results clearly indicate that recurrent neural network-based models, particularly the LSTM architecture, consistently outperform feedforward and convolutional models. This superior performance is attributed to their ability to effectively capture temporal dependencies and long-term patterns in PV power time series data. Compared with existing comparative studies, this work provides a unified and rigorous benchmarking framework in which all models are evaluated under identical datasets, forecasting settings, and validation criteria. The findings offer valuable practical insights into the selection and deployment of deep learning models for real-world PV power forecasting applications and support informed decision-making for grid operators and energy planners.