This study presents the development and evaluation of artificial neural networks (ANNs) and recurrent neural networks (RNNs) for photovoltaic power prediction using in-situ environmental data such as illumination, temperature, and humidity. Unlike conventional approaches based on machine learning libraries, the models were implemented entirely from scratch in Python version 3. This self-development of feedfordward propagation, error calculation, and backpropagation algorithms provided full control over the training process and promoted a deeper understanding of the mathematical principles behind ANNs. The developed ANN model achieved a coefficient of determination ( \(R^2\) ) of 0.97, confirming its strong predictive performance. The RNN model, designed to capture sequential dependencies, reached an \(R^2\) of 0.88. Both models incorporated early stopping and metric monitoring, which improved generalization and reduced overfitting. The low root mean squared error (RMSE) values obtained—across various combinations of sensor inputs—demonstrate the feasibility of accurate predictions even with limited environmental data.