Renewable Energy Sources generation forecasting provides short-term energy generation forecasts for different district RES assets for the next hours or days, with a granularity ranging from 15 minutes up to 1 hour in a time series format.
The RES generation forecasting module estimates renewable energy production within a PED, focusing initially on solar energy using PV data from the RTU campus. It consists of two main components:
- Solar Irradiance Prediction: A deep learning model (multilayer perceptron) forecasts short-term solar irradiance based on historical weather data (e.g. temperature, humidity, wind speed). Data is preprocessed using min-max scaling and outlier removal for improved accuracy.
- PV Energy Output Calculation: The predicted irradiance is used to calculate PV energy generation, factoring in panel characteristics such as tilt angle, efficiency, temperature coefficients, and surface area. A set of equations adjusts for local environmental conditions to estimate energy output over time.
This two-step approach enables accurate forecasting of solar energy production to support PED energy management.
The RES generation forecasting module provides hourly forecasts of renewable energy production for specific district assets, with a current focus on solar panel installations. Forecasts are generated for the next 24-hour period and are delivered in time series format through a REST API, enabling seamless integration with other digital services and energy management tools within the Positive Energy District (PED).
The forecasting methodology is based on historical data from three rooftop solar installations at the Riga Technical University (RTU) campus, covering the period from January 2023 to November 2024. This includes both PV output data and local weather observations from an RTU-operated station, such as temperature, humidity, wind speed, wind direction, and solar radiation. Additional environmental parameters - such as cloud cover and atmospheric pressure - were sourced from the Open-Meteo API to complete the dataset.
Using this combined dataset, a deep learning model is trained to predict short-term solar irradiance, which serves as the foundation for estimating energy output. The model - a multilayer perceptron (MLP) - takes historical weather variables as input and outputs predicted solar radiation (W/m²) for each time slot. Model performance indicators include a Mean Absolute Error (MAE) of 43.4 W/m², Mean Squared Error (MSE) of 6095.6, and a coefficient of determination (R²) of 0.684, demonstrating reliable performance across varying weather conditions.
Following irradiance prediction, the module applies a physics-based energy output calculation to estimate the actual electricity generation of the PV installations. This calculation incorporates metadata for each solar panel system, including tilt angle, surface area, conversion efficiency, nominal operating cell temperature, and temperature coefficient. Environmental conditions are also accounted for using parameters such as ambient temperature, clearness index, and reference temperature. The resulting energy forecasts reflect expected PV output under realistic, location-specific conditions.
| Data | Description | Type | Unit | Temporal resolution | Spatial resolution |
|---|---|---|---|---|---|
| renewable generation | metered energy generation of a RES asset | time series | kWh | 15 min or 1 hour | RES asset |
| outside temperature | temperature forecast from weather services | time series | °C | 1 hour | district |
| humidity | humidity forecast from weather services | time series | % | 1 hour | district |
| cloud cover | cloud cover forecast from weather services | time series | % | 1 hour | district |
| pressure | pressure forecast from weather services | time series | hPa | 1 hour | district |
| wind speed | wind speed forecast from weather services | time series | m/s | 1 hour | district |
| wind direction | wind direction forecast from weather services | time series | ° | 1 hour | district |
| PV area | area covered by PV panels for a RES asset | n/a | m² | n/a | RES asset |
| tilt angle | panel tilt angle for a RES asset | n/a | ° | n/a | RES asset |
| temperature coefficient | temperature coefficient for the solar cell for a RES asset | n/a | %/°C | n/a | RES asset |
| PV conversion efficiency | total conversion efficiency of the photovoltaic cell for a RES asset | n/a | % | n/a | RES asset |
| nominal operating cell temperature | Nominal Operating Cell Temperature for a RES asset | n/a | °C | n/a | RES asset |
| Data | Description | Type | Unit | Temporal resolution | Spatial resolution |
|---|---|---|---|---|---|
| renewable generation | energy generation forecast of a RES asset | time series | kWh | 15 min or 1 hour | RES asset |
Τhe module was developed using Python (https://www.python.org/). In terms of libraries and frameworks, TensorFlow (https://www.tensorflow.org/) and Keras (https://keras.io/) were utilised for the deep learning model. After training, the model is saved as Pickle files or H5 files. NumPy (https://numpy.org/) and Pandas (https://pandas.pydata.org/) are adopted for data processing, such as data loading, resampling, cleaning, and time series manipulation. Scikit-learn (https://scikit-learn.org/stable/) is used for data normalisation (e.g. min-max scaling) and outlier removal (e.g. z-score). The Open-Meteo API (https://open-meteo.com/) and the Norwegian Meteorological Institute API (https://www.yr.no/en) are used to retrieve historical and forecasted weather data when needed.