This article presents the study of a long-short-term memory (LSTM) neural network model to predict global solar irradiance (GHI) in the localities of Chaltura, Pimampiro, and Tumbabiro. The model’s lightweight architecture, with only 329 trainable parameters, allows for efficient training and mitigates the risk of overfitting, a particularly relevant aspect when working with moderate-sized datasets. The learning process shows rapid convergence, registering significant improvements from the fifth epoch onward, demonstrating the model’s ability to capture temporal patterns in historical data. The evaluation metrics used (root mean square error [RMSE], mean absolute error [MAE], and mean absolute percentage error [MAPE]) reflect high prediction accuracy. The RMSE values (between 2.023 and 2.769) indicate a slight deviation from observed values, while the MAE values (between 0.787 and 0.995) corroborate the model’s stability. Additionally, the low MAPE percentages (between 0.13% and 0.17%) highlight the model’s competitiveness compared to approaches applied to other renewable energy sources. The fit plots visually support the high fidelity of the predictions. Furthermore, cross-compression analysis between the time series reveals a maximum at zero lag, with coefficients greater than 0.6. This temporal synchronization, with no evidence of significant lags, suggests coupled solar dynamics between nearby locations, which favors the development of multivariate models without the need for additional temporal adjustments.