The methods described in these papers might help:

Article Fast Short-Term Global Solar Irradiance Forecasting with Wra...

Conference Paper Gaussian Process with Linear Discriminant Analysis for Predi...

Chapter Intra-hour Forecasting of Direct Normal Solar Irradiance Usi...

Article Minimum redundancy – Maximum relevance with extreme learning...

Hi Feddah,

You can use iterative cross-validation (important for small datasets):

https://otexts.com/fpp3/tscv.html

You can use recurrent neural nets for your model

Conference Paper Time Series Forecasting on Solar Irradiation using Deep Learning

If you have small datasets, you need models that have few parameters(low complexity) or strong prior.