ICIS - Power price prediction with neural networks
2 likes1,233 views
The document outlines the release of a new spot price forecast model for the German/Austrian power market, achieving a mean absolute error of 1.08€ for base price predictions. It explains neural network modeling techniques used in predicting power prices, including their architecture, learning process, and advantages compared to traditional models. Additionally, it discusses challenges in modeling power prices due to volatility and uncertainties in renewable generation and demand.
ICIS - Power price prediction with neural networks
- 1. Power Spot Price Prediction with Neural Networks Free webinar 12.05.2014 Jonathan Scelle Senior Analyst EU Power Markets Sebastian Stütz Lead Analyst Power
- 2. www.icis.com New Spot Price Forecast Model Our latest release for the German/Austrian power market Based on extensive research and academic cooperation Results as good as 2.35€ MAE for base (12 weeks) Today 1.08€ MAE base
- 3. www.icis.com New Spot Price Forecast Model – Today‘s Results 1.08€ MAE base, 2.61€ MAE hourly
- 4. www.icis.com €1.00 €2.00 €3.00 €4.00 €5.00 €6.00 €7.00 €8.00 7 8 9 10 11 12 13 14 15 16 17 18 Calendar Week Day Ahead 2 Day Ahead Base MAE per week Feb 10, 2014 – May 04,2014 1,42 € 3,78 € Ø MAEday-ahead: Ø MAE2-day-ahead:2.35 € 2.99 €
- 5. www.icis.com MAE per week €1.00 €2.00 €3.00 €4.00 €5.00 €6.00 €7.00 €8.00 7 8 9 10 11 12 13 14 15 16 17 18 Calendar Week Day Ahead 2 Day Ahead Feb 10, 2014 – May 04,2014 5,73 € 3,79 € Ø MAEday-ahead: Ø MAE2-day-ahead:4.34 € 4.66 € Due to negative prices on the 16 th of March
- 6. www.icis.com March 15, 2014 - March 16, 2014 Ø MAE March 16th: 12.93 €
- 7. www.icis.com Content PART I – Background in Neural Network Modeling PART II – New Model for Power Spot Markets Your Questions
- 8. www.icis.com What Are Neural Networks? The main concept of NN originated in the 1940‘s Idea: Build a simplified model of the brain to get an universal function approximator Initially NNs had very little practical use due to computational limitations Because of technological advances in recent years, NNs can now be used for a wide variety of tasks in many different business fields NN Inputs Inputs Outputs Outputs
- 9. www.icis.com Why Use Neural Networks? NNs can learn from sample data NNs are data driven self-adaptive models which determine their function based on sample data No a-priori assumptions are needed NNs can generalize NNs can produce reasonable outputs for previously unseen data NNs are universal function approximators NNs can deal with non-linear relationships NNs are successfully used for a wide variety of tasks Facial Recognition Text analysis Technical process control Medical diagnosis Stock market forecasts
- 10. www.icis.com How Do Neural Networks Work? (1/2) - Architecture 1. Hidden LayerInput Layer Output Layer Input Neuron Hidden Neuron Output Neuron 2. Hidden Layer An NN consists of one input layer, one output layer and any desired amount of hidden layers Each layer consists of one or more neurons Most real world problems require at least one hidden layer
- 11. www.icis.com How Do Neural Networks Work? (2/2) - Neurons The output of a neuron is a function of the weighted sum of the inputs The function of the entire Neural Network is the computation of the outputs of all the neurons An entirely deterministic calculation ƒ Fact Output = (in1* w1 + in2* w2+ in3* w3) in1 in2 in3 Fact w1 w2 w3 Activation Function ∑ Neuron
- 12. www.icis.com How Do Neural Networks Learn? (1/5) Goal: Minimize the discrepancy between real data and the output of the network Discrepancy between the target values and the output of the network is measured with an error function e.g. RMSE How: By adjusting the weights associated with the connections between the neurons
- 13. www.icis.com How Do Neural Networks Learn? (2/5) We are looking for the minimum of the error function in weight space Problem: No mathematical method exists that guarantees to find the global minimum W1 W2 E Source: Oberhofer, W., T. Zimmerer, and D.-K. T. Zimmerer (1996). Wie Künstliche Neuronale Netze lernen, p. 16
- 14. www.icis.com How Do Neural Networks Learn? (3/5) Backpropagation - Algorithm: Calculation of output based on current weights as initialization Method of gradient descent Iterative method – in each iteration the weights are modified Performance highly depends on two parameters: Learning rate: Determines the size of the weight changes Momentum: Determines how past weight changes affect current weight changes Learning rate and momentum can only be optimized by a trial and error process
- 15. www.icis.com How Do Neural Networks Learn? (4/5) Depending on the starting weights, the Backpropagation – Algorithm can converge to different minimums of the error function Different random sets of starting weights can improve the chances of reaching a global minimum W1 W2 E W1 W2 E Source: Oberhofer, W., T. Zimmerer, and D.-K. T. Zimmerer (1996). Wie Künstliche Neuronale Netze lernen, p. 16
- 16. www.icis.com How Do Neural Networks Learn? (5/5) Implications: The shape of the error function is influenced by The number of connections of a NN The activation function of the neurons The inputs of a NN The shape of the error function influences the optimal values of the learning rate and the momentum The learning rate and the momentum influence the performance of the Backpropagation Algorithm and thus the quality of the NN’s predictions For each set up of a NN the learning rate and the momentum have to be optimized (computationally very expensive)
- 17. www.icis.com How Long Should A Neural Network Be Trained? The Problem of Overfitting: NN learns random errors or noise in the sample data instead of underlying relationships Causes: Too many parameters (connections) relative to the training set size (model is too complex) Consequences: Poor prediction results on unseen data Solution – Early Stopping: Stop training if error of validation set starts to increase Error Iterations Training Set Validation Set Stop training
- 18. www.icis.com Designing a Neural Network Input Identification Data Pre- Processing Input Decision Neural Network Optimiziation (Parameters) Final Training Production Model
- 19. www.icis.com Oberhofer, W., T. Zimmerer, and D.-K. T. Zimmerer (1996). Wie Künstliche Neuronale Netze lernen: Ein Blick in die Black Box der Backpropagation Netzwerke. Univ., Wirtschaftswiss. Fak. Crone, S. F. and N. Kourentzes (2010). Feature selection for time series prediction–a combined filter and wrapper approach for neural networks. Neurocomputing 73 (10), 1923–1936. Rojas, R. (1996). Neutral Networks: A Systematic Introduction. Springer. Prechelt, L. (1998). Automatic early stopping using cross validation: quantifying the criteria. Neural Networks 11 (4), 761–767. Kaastra, I. and M. Boyd (1996). Designing a neural network for forecasting financial and economic time series. Neurocomputing 10 (3), 215–236. References
- 20. www.icis.com Content PART I – Background in Neural Network Modeling PART II – New Model for Power Spot Markets Your Questions
- 21. www.icis.com How to model power prices? Market Challenges Uncertainty in renewable gener- ation and power demand High day to day price volatility Negative prices
- 22. www.icis.com How to model power prices? Which prices to forecast? Spot vs. forward market (spot market highly volatile with renewable challenges) OTC, daily auction Specific power market problems to address State of information before auction gate closure How to model hours? Separate prediction / 24h prediction? (number of inputs in model / complexity / overfitting?) Multiple day ahead => feed same model / new model? Negative prices
- 23. www.icis.com Model Key Summary Please contact us for a detailed technical specification of the model * Import / Export explicit modeling is running project.
- 24. www.icis.com Implicit stack / merit order approximation Our model is not trained to forecast absolute prices but to learn price gradients in the merit order, visible through auction results (extension to bidding curves in plan) The trained “model” can be described as an experienced view on price gradients at the price setting parts of the merit order Hence, the model is capable of predicting price changes from drops/increases in e.g. residual load or available capacities In order to distil changes in historic data we normalize always based on each latest week. Our running forecasts take into account latest days and weeks and long-term trends. Source: Risø DTU
- 25. www.icis.com Advantages / Disadvantages of Neural Networks as Power Price Models Pros Decreases need for explicit assumptions How do you model in your stack… Actual efficiencies and capacities of each plant? Inland transportation costs Topping turbines Combined heat and process steam generation Must run conditions Transferable to other markets Constantly learning Cons Require long series, structural change of market mechanisms (e.g. capacity market) would be a problem Computationally expensive
- 26. www.icis.com Outlook – running ICIS Analytics projects Explicit Import / Export Modeling, based on border balances with e.g. Temperature driven power demand deviations from norm Holidays and seasonal patterns in public behavior Renewable forecasts France EPEX spot auction price model Dedicated negative price model
- 27. More power data for market professionals ICIS ANALYTICS POWER CONTENT Along with our spot price forecasts, the ICIS Power Portal provides you with: POWER FORECASTS • Neural Network based power demand forecasts (DE/AT only) • Renewable generation forecasts wind and solar for nearly all European markets, comparison to norm • Forecasts for power demand changes due to temperatures, comparison to norm • Forecasts for precipitation, translated into power production potential NEWS & PRICES • Intra-day updates for power news on all major European markets • Latest and historic OTC power prices for all major European markets FUNDAMENTAL DATA • Access to power relevant data sources all in one place (e.g. ENTSOE, TSOs, EEX) BEHAVIOUR DATA • Tracking of power forward hedging strategies of major European utilities
- 28. www.icis-analytics.com/ Request a free trial now. Jonathan Scelle Senior Analyst EU Power Markets Jonathan.scelle@icis.com Sebastian Stütz Lead Analyst Power Sebastian.stuetz@icis.com