The electricity market is driven by complicated interactions that are hard to model analytically. This is particularly the case for the balancing market, where imbalances between supply and demand after the day-ahead market clearance are balanced. The balancing market bridges the gap between the dayahead market and the actual power system operations. Being able to predict the necessary balancing volumes and prices some hours in advance of the operational hour will allow power producers to plan their production and trading in a more optimal way. There exist large amounts of open data that could contain predictive information about the balancing market, including day-ahead market data and climatic data. However, the literature on forecasting volume and prices in the balancing market is sparse compared to the rich literature on forecasting for the day-ahead market. Neural networks are powerful functional approximators and well-suited to model the complex relationships in the power market. It may also be used to study the predictability of the balancing volumes and prices forward in time. In this paper, we develop a model based on long short-term memory (LSTM) recurrent neural networks to predict volumes and prices in the Nordic balancing market based on public accessible data. Results show that the LSTM model performs well when compared to the two baselines selected. However, the performance is not significantly better, which indicates that the market data does not hold significant predictive information.

No abstract

Videos and images are commonly used in home monitoring systems. However, detecting emergencies in-home while preserving privacy is a challenging task concerning Human Activity Recognition (HAR). In recent years, HAR combined with deep learning has drawn much attention from the general public. Besides that, relying entirely on a single sensor modality is not promising. In this paper, depth images and radar presence data were used to investigate if such sensor data can tackle the challenge of a system's ability to detect abnormal and normal situations while preserving privacy. The recurrence plots and wavelet transformations were used to make a twodimensional representation of the presence radar data. Moreover, we fused data from both sensors using data-level, feature-level, and decision-level fusions. The decision-level fusion showed its superiority over the other two techniques. For the decision-level fusion, a combination of the depth images and presence data recurrence plots trained first on convolutional neural networks (CNN). The output was fed into support vector machines, which yielded the best accuracy of 99.98%.