Rising global energy demand, predominantly satisfied by fossil fuels, triggers fuel price surges, fuel scarcity, and substantial greenhouse gas emissions. Solar photovoltaics (PV), as an abundant renewable alternative, can potentially address this demand, yet low cell efficiency (15-25%) and fluctuating output power due to intermittent irradiance ( G ) and temperature ( T ) impedes grid integration. This paper presents a novel Deep Learning (DL) based stacked LSTM (Long Short-Term Memory) MPPT controller to maximize power harvesting from a 100 kW grid-tied solar PV system, demonstrating superiority over conventional Perturb & Observe (P&O) and Feed Forward-Deep Neural Network (FF-DNN) MPPT approaches. Subsequently, a Neutral-Point-Clamped (NPC) 3-level inverter with proportional-integral (PI) controllers regulates the DC link voltage and transfers the extracted PV power to the grid. The proposed MPPT methodology includes collection of one million-sample ( G , V , Vmp ) datasets; preprocessing via z-score normalization; visualizing distributions through histograms and correlation matrix plots; an 80/20 split rule-based training and test sets; a two-hidden layer stacked LSTM (64 and 32 neurons) architecture; hyperparameters including the Adam optimizer, 0.05 learning rate, 32 batch size, and 50 epochs. Model efficacy quantification uses MSE, RMSE, MAE, loss, and R2 metrics. For 100 kW generated PV power, the stacked LSTM extracts 98.2 kW, versus 96.1 kW and 94.3 kW for the DNN and P&O MPPTs respectively. By integrating the optimized proposed stack LSTM MPPT with a streamlined inverter topology, the proposed approach advances the state-of-the-art in DL based solar PV energy harvesting optimization and grid integration.