Probal Roy

Probal Roy

Auto-generated profile, based on publicly available information

Probal Roy

Universiti Malaysia Sarawak

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
14 October 2026 in VFAST Transactions on Software Engineering
Verified
Journal Article
Open access
PDF file
PRProbal Roy
Probal Roy
IRIr. Dr. Andrew Ragai Anak Henry Rigit
Ir. Dr. Andrew Ragai Anak Henry Rigit
UMUmair Mukhtiar
Umair Mukhtiar
AJAhmad Fadzil Jobli
Ahmad Fadzil Jobli
SKSarken Kapayeva
Sarken Kapayeva
AVAnatoli Vakhguelt
Anatoli Vakhguelt
Solar–hydro hybrid energy systems combine rapidly varying photovoltaic output with slower hydraulic dynamics, necessitating source-aware online adaptation. This study evaluates a software-in-the-loop monitoring and forecasting framework using a compact Hedge-adaptive backpropagation neural network with separate solar and hydro feedback gains. The dataset comprises 3,843 mixed-source observations collected from 18 July 2026 at 22:35 to 17 August 2026 at 00:13. Irradiance and ambient temperature were obtained from external weather-model outputs, rainfall was retrieved from a public service when available, hydraulic states were simulated, and hybrid power was constructed as a reference rather than obtained from independently metered plant output. Feature engineering generated 3,823 model-ready targets, which were evaluated using a chronological 60/20/20 development–validation–test split, with scaling fitted exclusively on the development set. On 764 common final-test targets, dual feedback reduced the mean absolute percentage error (MAPE) from 6.80 ± 3.04% for the matched no-feedback Hedge-BPNNT to 2.47 ± 0.57% across five prespecified random seeds. However, support vector regression (SVR) achieved the lowest one-step MAPE of 0.96% and the lowest MAPE at all 24 continuity-safe forecast lead times in the four-hour comparison. For the largest 10% of ten-minute power changes, dual feedback outperformed persistence, long short-term memory (LSTM), and the no-feedback variant, but remained inferior to SVR and Random Forest. The mean computation time for dual feedback was 0.111 ms per one-step prediction and update and 3.409 ms per 24-step recursive forecast. These findings indicate that source-specific feedback provides a useful adaptation mechanism within a matched neural-network architecture, while SVR delivers the strongest overall forecasting accuracy on the constructed-reference dataset.