Baptiste Schubnel

Accurate day-ahead forecasts of solar irradiance are required for the large-scale integration of solar photovoltaic (PV) systems into the power grid. However, current forecasting solutions lack the temporal and spatial resolution required by system operators. In this paper, we introduce SolarCrossFormer, a novel deep learning model for day-ahead irradiance forecasting, that combines satellite images and time series from a ground-based network of meteorological stations. SolarCrossFormer uses novel graph neural networks to exploit the inter- and intra-modal correlations of the input data and improve the accuracy and resolution of the forecasts. It generates probabilistic forecasts for any location in Switzerland with a 15-minute resolution for horizons up to 24 hours ahead. One of the key advantages of SolarCrossFormer its robustness in real life operations. It can incorporate new time-series data without retraining the model and, additionally, it can produce forecasts for locations without input data by using only their coordinates. Experimental results over a dataset of one year and 127 locations across Switzerland show that SolarCrossFormer yield a normalized mean absolute error of 6.1 % over the forecasting horizon. The results are competitive with those achieved by a commercial numerical weather prediction service.

Accurate forecasting of solar power generation with fine temporal and spatial resolution is vital for the operation of the power grid. However, state-of-the-art approaches that combine machinelearning with numerical weather predictions (NWP) have coarse resolution. In this paper, we take a graph signal processing perspective and model multi-site photovoltaic (PV) production time series as signals on a graph to capture their spatio-temporal dependencies and achieve higher spatial and temporal resolution forecasts. We present two novel graph neural network models for deterministic multi-site PV forecasting dubbed the graph-convolutional long short term memory (GCLSTM) and the graph-convolutional transformer (GCTrafo) models. These methods rely solely on production data and exploit the intuition that PV systems provide a dense network of virtual weather stations. The proposed methods were evaluated in two data sets for an entire year: 1) production data from 304 real PV systems, and 2) simulated production of 1000 PV systems, both distributed over Switzerland. The proposed models outperform state-of-the-art multi-site forecasting methods for prediction horizons of six hours ahead. Furthermore, the proposed models outperform state-of-the-art single-site methods with NWP as inputs on horizons up to four hours ahead.

We present a three-step method to perform system identification and optimal control of nonlinear systems. Our approach is mainly data-driven and does not require active excitation of the system to perform system identification. In particular, it is designed for systems for which only historical data under closed-loop control are available and where historical control commands exhibit low variability. In the first step, simple simulation models of the system are built and run under various conditions. In the second step, a neural network architecture is extensively trained on the simulation outputs to learn the system physics and retrained with historical data from the real system with stopping rules. These constraints avoid overfitting that arises by fitting closed-loop controlled systems. By doing so, we obtain one (or many) system model(s), represented by this architecture, whose behavior can be chosen to match more or less the real system. Finally, state-of-the-art reinforcement learning with a variant of domain randomization and distributed learning is used for optimal control of the system. We first illustrate the model identification strategy with a simple example, the pendulum with external torque. We then apply our method to model and optimize the control of a large building facility located in Switzerland. Simulation results demonstrate that this approach generates stable functional controllers that outperform on comfort and energy benchmark rule-based controllers.

We introduce the Python-based open-source library Energym, a building model library to test and benchmark building controllers. The incorporated building models are presented with a brief explanation of their function, location and technical equipment. Furthermore, the library structure is described, highlighting the necessary features to provide the benchmarking and control capabilities, i.e., standardized evaluation scenarios, key performance indicators (KPIs) and forecasts of uncertain variables. We go on to characterize the evaluation scenarios for each of the models and give formal definitions of the KPIs. We describe the calibration methodologies used for constructing the models and illustrate their usage through examples.

Operating power systems with large amounts of renewables requires predicting future photovoltaic (PV) production with fine temporal and spatial resolution. State-of-the-art techniques combine numerical weather predictions with statistical post-processing, but their resolution is too coarse for applications such as local congestion management. In this paper we introduce computing methods for multi-site PV forecasting, which exploit the intuition that PV systems provide a dense network of simple weather stations. These methods rely entirely on production data and address the real-life challenges that come with them, such as noise and gaps. Our approach builds on graph signal processing for signal reconstruction and for forecasting with a linear, spatio-temporal autoregressive (ST-AR) model. It also introduces a data-driven clear-sky production estimation for normalization. The proposed framework was evaluated over one year on both 303 real PV systems under commercial monitoring across Switzerland, and 1000 simulated ones based on high-resolution weather data. The results demonstrate the performance and robustness of the approach: with gaps of four hours on average in the input data, the average daytime NRMSE over a six-hour forecasting horizon (in 15 min steps) and over all systems is 13.8% and 9% for the real and synthetic data sets, respectively.

Power consumption in buildings show nonlinear behaviours that linear models cannot capture, whereas recurrent neural networks (RNNs) can. This ability makes RNNs attractive alternatives for the model-predictive control (MPC) of buildings. However, RNNs are nonlinear and non-smooth functions which makes their use challenging in optimization problems. Therefore, this work systematically investigates whether using RNNs for building control provides net gains in MPC. It compares over 2 months of simulated operation the representation power and control performance of two architectures: an RNN architecture and a linear state-space (LSS) model with a nonlinear regressor to estimate energy consumption. The results show that RNNs yield an identification error 69% lower than LSS, but the LSS models yield control laws that achieve 10% lower objective function with a computational time three times lower than the RNNs. Thus, on balance, well-designed LSS models with nonlinear regressors are best in most cases of MPC.

The electricity sector foresees a significant change in the way energy is generated and distributed in the coming years. With the increasing penetration of renewable energy sources, smart algorithms can determine the difference about how and when energy is produced or consumed by residential districts. However, managing and implementing energy demand response, in particular energy flexibility activations, in real case studies still presents issues to be solved. This study, within the framework of the European project “SABINA H2020”, addresses the development of a multi-level optimization algorithm that has been tested in a semi-virtual real-time configuration. Results from a two-day test show the potential of building’s flexibility and highlight its complexity. Results show how the first level algorithm goal to reduce the energy injected to the grid is accomplished as well as the energy consumption shift from nighttime to daytime hours. As conclusion, the study demonstrates the feasibility of such kind of configurations and puts the basis for real test site implementation.

Here is information on how to publish, spread your research, and search for your publications. Most of LiU's research publications are directly imported into DiVA from the Web of Science. Before registering a publication, always check if the publication is already in DiVA!