Many 19thcentury Romanian texts are written in the Transitional Script (RTS) combining Cyrillic and Latin characters in variable proportions depending on specific factors (period, region, publishing house, literary trends, and personal beliefs). Thousands of heritage documents in numerous libraries across Romania have yet to be digitized. Reading texts written in RTS is challenging for modern scholars, today’s students and OCR engines which due to specific particularities have difficulties in processing them. Transfer learning can be used to train state-of-the-art tools such as Tesseract or Transkribus, but their efficiency varies according to the underlying deep neural network, the training parameters, and the particularities of the scanned images and used scripts. In this paper, we compare the performance of two architectures on 180 OCR models for 5 scenarios (best CER <4.6% for Tesseract) and discuss OCR challenges on RTS texts.

We use satellite imagery to identify cloud types and correlate with irradiance data from a solar platform. Processing the large images takes several minutes. Integrating into a web service would enable multiple users to estimate solar irradiance using just satellite imagery. It would also allow for the processing of many images at once as more remote sensing satellites are launched and data is easily accessible. The Sentinel 2 Cirrus Band shows an 81 % correlation with the diffuse solar irradiance values measured at the platform.

Nowcasting is vital for PV farms in order to match supply and demand on energy markets. The accuracy of existing forecasting methods relies on short term modelling of cloud movement and dynamics by combining satellite images with ground based allsky observations. Cloud movement is modelled based on complex nonlinear numerical models which on the short term are outperformed by image analysis techniques. In this research we propose a simplified nature inspired model for nowcasting which scales with the data. The model is based on forecasting wind direction extracted from motion vector fields by modelling wind using a modified Boids Flocking algorithm. We show that the model is accurate in predicting the cloud coverage up to several hours ahead and that it scales with the number of particles on shared memory systems suited for commodity machines available to PV farm operators.