Industry
End-to-end production platform using machine learning for site-specific solar-generation and facility-load forecasts, refreshed every 15 minutes over a three-day horizon and integrated with downstream energy-management workflows.
Prepared historical and incoming data used for solar-generation and facility-load forecasting, including site measurements, weather information, irradiance, and calendar features.
The processing workflow aligns these heterogeneous time series on a consistent temporal basis and prepares them for both model training and operational inference.
Validation steps were introduced to identify incomplete, stale, or inconsistent inputs before they reached the forecasting models.
The same preparation logic was used across site-specific forecasting problems, providing a consistent input layer for training, backtesting, and forecast generation.
The platform produces separate forecasts for solar generation and facility load, with models configured for individual sites and measurement types.
Forecasts are refreshed every 15 minutes and cover a three-day horizon at the same temporal resolution.
Forecast outputs are made available to downstream energy-management workflows, where expected generation and consumption can be incorporated into the operation of distributed energy resources.
The forecasting layer was designed as part of a wider operational system rather than as a standalone modeling experiment.
Developed forecasting models separately for each site and forecast target, allowing the model to adapt to local solar-production and facility-load patterns.
The modeling framework included gradient-boosted models, linear models, and simple statistical baselines.
Candidate models were evaluated on historical periods before deployment, with forecast performance compared across sites and forecast horizons using standard error metrics.
This provided a reproducible way to select an appropriate model for each forecasting problem rather than applying a single model architecture everywhere.
Models could be retrained as additional operational data became available or site behavior changed, while retaining the same forecasting and evaluation workflow.
Developed monitoring tools to compare forecasts with measured solar generation and facility load across sites and forecast horizons.
The monitoring workflow made it possible to review forecast errors over time and identify changes in model performance.
Input-data health and forecast execution were monitored alongside model accuracy, helping distinguish forecasting errors from missing, stale, or otherwise problematic upstream data.
Interactive visualizations supported detailed inspection of individual sites as well as comparison across the forecasting portfolio.
Site and model configurations were separated from the core forecasting logic so that the same platform could support different solar-generation and facility-load forecasting problems.
This approach reduced site-specific implementation work while keeping individual forecasting configurations independently manageable.
The work covered model development through operational deployment, including configuration, forecast generation, monitoring, model updates, and integration with the surrounding energy-management platform.
The platform was used for operational forecasting rather than being limited to offline research or experimentation.
End-to-end development and modernization of a data platform spanning production ingestion, analytics engineering, quality controls, and reporting for operational, business, and customer-facing use cases.
Spatiotemporal forecasting of regional photovoltaic production from Meteosat image sequences using a ConvLSTM model trained and evaluated against measured production and a persistence baseline.
Scientific framework for evaluating solar forecasts beyond average error metrics, using Temporal Distortion Mix and a ramp metric to assess timing errors and rapid irradiance variations.
