Time series problems in the energy industry
Time series modelling is omnipresent in the energy industry. The modeling problems include electricity load forecasting, prediction of technical losses in the grid or distribution network, calculation of system imbalances, gas consumption, wind and solar production or electricity price forecasting.
These challenges can be represented as a time series where historical measurements of the target variable are explained by a variety of predictor candidates or features derived out of them. In the energy industry, the predictor candidates are often weather and climate related variables, calendar information etc. The features often include time delays of original predictors, their interactions, moving averages and others.
TIM, Tangent Information Modeler, is a modeling engine that automates the process of feature engineering, model building and forecasting allowing for large scale automatic time series modeling.
The data science process
We use a picture to describe the data science process and explain the role TIM plays in it.
Steps 1 to 3 are related to getting good data and preprocessing. The quality of data has an essential impact on forecasting accuracy. The value of TIM lies in step 4 and 5, the actual building of the model and calculating the forecast for the required scenario.
Model building consists of two main tasks, feature engineering (FE) and model selection (MS). The crucial questions in the FE process are “Which features and how many?”. FE is therefore often a tedious task as it requires a search through a vast number of possible features.
The number of potential features s grows exponentially with the number of predictors resulting in much time and expertise required for the engineering.
The selected features are then used in MS process where an expert selects a modelling technique while tuning its meta-parameters, i.e. selecting a neural net with “m” hidden layer of “n” units over some other configuration or modelling technique.
Why TIM builds models differently?
Unlike other popular techniques such as Neural Networks, Deep Learning algorithms and Support Vector Machines, TIM does not create black box models. TIM’s models are a transparent formula with interpretable terms. This allows us to see into the dynamics of underlying data.
The process has the following steps:
- Data intake
- Feature expansion
- Feature synthesis, based on heuristic rules
- Feature selection, carried out in Euclidean space using Bayesian inference
- Model Complexity is estimated using a customized BIC
- Model is generated
Traditional methods require feature engineering to allow the model to find stronger relationships between parameters and increase the model’s predictive power. TIM does the feature engineering and model building automatically in seconds without requiring costly cross validation of your models.
In the classical approach to time series modeling, parameter estimates (weights) are optimized for each time series and for each model candidate. This means that several optimizations per time series. This takes time, expertise and processing power. TIM eliminates this problem. TIM generates solid models with high standard of predictive power without model tuning.
TIM builds non-parametric non-linear models directly from the data. The model building is driven by information criteria that optimize generalization directly, rather than fitting a single model to the data. Hence the models that TIM generates have excellent generalization and predictive power in an operational context.
How accurate is TIM?
TIM has clients whose businesses rely on TIM as a large-scale forecasting engine to drive their operations. TIM also proved its accuracy by winning GEFCOM 2017.
The Global Energy Forecasting Competition was a way to test TIM’s effectiveness. GEFCOM attracts skilled data science teams from around the world, provides them with the same time series data, and then compares their forecasting results. Forecasts by TIM managed to beat other teams and won the competition.
When interpreting the results, please keep in mind that all of TIM results were forecasted automatically and in seconds, not requiring technical and mathematical know-how.
TIM allows for consistent, highly accurate model building and forecasting. TIM helps you as an expert to focus on what matters, creating the insights and business relevant conclusion while TIM helps you with the model building and forecasting itself.
TIM helps you to scale within the organization so you and your team can cope with the ever-increasing demand for forecasts in the industry.
Watch a Demo of TIM
View the video below to see TIM in action. It shows TIM’s place in your data workflow, the model generation process and the resulting forecasts.