Context

This open dataset is used in the scientific study "Coordination of operational planning and real-time optimization in microgrids" currently in the submission process of the Power Systems Computation Conference (PSCC) 2020.

The paper abstract is:

Hierarchical microgrid control levels range from distributed device level controllers that run at a high frequency to centralized controllers optimizing market integration that run much less frequently. Centralized controllers are often subdivided in operational planning controllers that optimize decisions over a time horizon of one or several days, and real-time optimization controllers that deal with actions in the current market period. The coordination of these levels is of paramount importance. In this paper we propose a value function based approach as a way to propagate information from operational planning to real-time optimization. We apply this method to an environment where operational planning, using day-ahead forecasts, optimizes at a market period resolution the decisions to minimize the total energy cost and revenues, the peak consumption and injection related costs, and plans for reserve requirements, while real-time optimization copes with the forecast errors and yields implementable actions based on real-time measurements. The approach is compared to a rule-based controller on three use cases, and its sensitivity to forecast error is assessed.

Content

The dataset is composed:

• 15 minutes resolution weather forecast (solar irradiation, air temperature, etc) produced by the Laboratory of Climatology of the university of Liège, based on the MAR regional climate model;
• 5 seconds resolution monitored PV and consumption of the MiRIS microgrid;
• 15 minutes resolution PV and consumption weather based forecasts produced by our forecaster and used in the study.

The weather based forecasts are multi outputs with an horizon of 24 hours ahead and a resolution of 15 minutes. They are quarterly produced on rolling basis with a learning set of one week. Every six hours the model is refreshed and the learning set is moved consequently. This means, each quarter a PV and consumption forecast is produced, composed of 96 values (one per quarter of the 24 hours ahead).

More information about the MiRIS microgrid located at the John Cockerill Group’s international headquarters in Seraing, Belgium, is available at https://johncockerill.com/fr/energy/stockage-denergie/.

Acknowledgements

We would like to thank John Cockerill and Nethys for their financial support, and Xavier Fettweis of the Laboratory of Climatology of ULiège who produced the weather forecasts based on the MAR regional climate model.

Inspiration

You can freely use the data to reproduce the numerical results of the study or to produce better weather based forecasts.

Weather based forecast description

Two "classic" deterministic techniques are implemented, a Recurrent Neural Network(RNN) with the keras python library and a Gradient Boosting Regression (GBR) with the scikit-learn python library.

The RNN is a LSTM with one hidden layer, 5000 epochs, RELU as activation function (hidden layer and output), a batch size of 200 and drop out rate of 0.4.

The GBR is the multi output GBR of sklearn with 200 estimator and 20 as max depth.

Weather forecast description

The weather forecast is made at the MiRIS microgrid loation.

The file is composed of forecast of several weather variables:

• CD = low clouds (0 to 1)
• CM = medium clouds (0 to 1)
• CU = high clouds (0 to 1)
• PREC = precipitation (mm / 15 min)
• RH2m = relative humidity (%)
• SNOW = snow height (mm)
• ST = Surface Temperature (°C)
• SWD = Global Horizontal Irradiance (W/m2)
• SWDtop = Total Solar Irradiance at the top of the atmosphere (W/m2)
• TT2M = temperature 2 meters above the ground (°C)
• WS100m = Wind speed at 100m from the ground (m/s)
• WS10m =Wind speed at 10m from the ground (m/s)