Data-driven learning from dynamic pricing data - Classification and forecasting

Inspired by recent advances in data driven methods from deep-learning, this paper shows how neural networks can be trained to extract valuable information from smart meter data. We show how these methods can help provide new insight into the effectiveness of dynamic time of use pricing schemes. In addition we show how long-short term memory networks, a particular form of recurrent neural networks, allows including the information of dynamic prices to improve the accuracy of load forecasting. The renewables transition require flexibility sources to replace the regulation capability of traditional generation. Buildings have a large capacity to supply part of this flexibility by adjusting their consumption taking into account the needs of the energy systems. The use of time-of-use pricing is one of the simplest form of demand side management, but the effectiveness of such schemes are often hard to quantity. The smart meter roll-out is expected to help provide bring about new understanding of consumption patterns - but methods to analyse the data and extract the relevant information are needed. The energy domain is still relying on methods for data analysis that are time consuming, does not scale and require costly manual handling. The methods demonstrated learn from real data from a trial with dynamic time-of-use pricing in London, UK.