Solar irradiance is a key energy input for photovoltaic power generation systems, and its forecasting accuracy directly affects photovoltaic power prediction, grid-connected dispatch, energy storage optimization, and operational decision-making. To address the nonlinear, non-stationary, and multi-scale fluctuation characteristics of solar irradiance series caused by diurnal cycles, seasonal variations, and complex meteorological disturbances, this study proposes a hybrid forecasting framework integrating an improved triangular topology aggregation optimizer (ITTAO) with temporal convolutional network (TCN), bidirectional gated recurrent unit (Bi-GRU), and attention mechanism. Based on the original triangular topology aggregation optimizer, the proposed framework introduces adaptive weight adjustment, differential evolution perturbation, and triangular topology collaborative updating mechanisms to enhance global exploration, local exploitation, and optimization stability during hyperparameter search. Meanwhile, TCN is used to extract local multi-scale fluctuation features, Bi-GRU is employed to model bidirectional temporal dependencies, and the attention mechanism is introduced to emphasize key time steps and meteorological features. Using two public photovoltaic station datasets, Clark and Loyola, monthly, seasonal, and short-term multi-step forecasting experiments at 1 h, 3 h, 6 h, and 24 h horizons were conducted. The proposed framework was compared with traditional machine learning models, deep learning models, and time-series baselines implemented under a unified experimental framework, including PatchTST-style and iTransformer-style models. The results show that the proposed framework achieves high accuracy in monthly and seasonal forecasting, with the lowest mean absolute error of 3.25 W·m?2 and the highest coefficient of determination of 0.998. In short-term multi-step forecasting tasks, model performance exhibits clear site- and horizon-dependent characteristics, while ITTAO-TBA reduces RMSE compared with the original TBA in several short- and medium-term forecasting tasks, indicating that ITTAO has a positive effect on TBA hyperparameter optimization. This study provides a competitive solar irradiance forecasting method for day-ahead and intraday photovoltaic dispatching and system optimization.