A world at the crossroads

1. The weather

How to determine in advance the  market price of a financial product? While the answer remains a martingale in most cases, the power market seems to stand out. Although there is no exception to the rule of supply-demand equilibrium, its specificity lies in the inelasticity of demand to the market price: you and I do not consume according to this price. This particularity neutralises part of the psychological randomness of the players and makes price modelling possible.However, there is indeed a significant amount of variability that lies at the crossroads of the energy and weather sectors. Focussing on electricity consumption, France is a reference to illustrate the concept of thermo-sensitivity: 1°C less in winter leads to an increase of 2,300 MW for consumption (the equivalent of two nuclear reactors) in France compared with 5,000 MW throughout Europe. Regarding supply and in a context of energy transition, the anticipation of production from renewable energies remains largely dependent on weather conditions. As electricity can not be stored – at least not in sufficient quantities for a realistic price – and demand must be met “at all costs”, the intermittent nature of these production methods dictates the market price. It is therefore understandable that the quality of the price forecast will depend first and foremost on the best possible forecast of its weather-sensitive components.

2. The job(s)

If it then seems possible to predict the electricity price thanks to, in particular,  weather forecasts, the solutions provided are conditioned by the identification of the needs of the various market players. Traditionally, three types of players are active on the market: (i) producers who trade and sell the output of their power plants, (ii) traders/dealers who wish to take advantage of variations to maximise their profits in line with their risk parameters, and (iii) suppliers who must anticipate their customers’ future consumption and secure, over the duration of the contracts, a supply at a cost consistent with their margin objectives. In addition, new players are emerging, including electric vehicle recharging networks, erasing operators – enhancing consumption flexibility – and micro-grids aggregating end-users. Every day, these players, both industrial and financial, must therefore position themselves in relation to their expectations of electricity price trends in the short and medium term and arbitrate purchases and sales.

3. The data

Given the large amount of data used, Machine Learning techniques are at the heart of the decision support tools implemented to best meet these needs.
Each of these players may have different objectives: search for a minimum absolute error, comparison with the OTC market, search for specialized models in very specific situations. Therefore, the first mission, which consists in converting this  need into a mathematical problem, is certainly one of the most important. Hence the need to understand what is at stake in this singular market. Followed by the classical steps of the heart of the data scientist’s job, including the recovery and cleaning of relevant data (without letting oneself be too influenced by one’s a priori to avoid any human bias), the selection of part of the spectrum offered by artificial intelligence (regression, classification, clustering), the optimisation of mathematical models, historical simulations, and performance evaluation validating or not its production. Maintenance, quality monitoring and product improvement come to complete the picture.

After bringing this trio of experts together, but also after giving them a certain amount of time to consider, to emulate, to fail and to succeed, experience shows that it is possible to achieve the much hoped-for model of artificial intelligence. The challenge is then to make it operational in real time in order to provide what is certainly of the greatest value: high-performance forecasts delivered quickly. It is up to you to define the “efficient” and the “fast”. Still, there is a high probability that your trio has selected input data for your model that is as diverse in format as it is in origin. More than hoping for consistency in this heterogeneity in order to not disrupt your processes, the necessary (but not sufficient) condition to avoid unpleasant surprises is to systematically check the integrity of this input data. Finally, despite all efforts and because there are always exceptional situations, errors will always remain. The trick is to know how to analyse them and adapt accordingly. Being able to recognize when you make a mistake is commendable, but let’s not forget to (humbly) take advantage of the moments when the work pays off.

Thomas Gossot
Data Scientist

COR-e Logo

1 rue Hoche

83000 Toulon, France