IWoRE, La Serena, 5 th July 2018 Renewable Energy research for Chile and Germany Dr. Martin Felder, Anton Kaifel, Kay Ohnmeiß, Dr. Frank Sehnke, Leon Schröder Zentrum für Sonnenenergie- und Wasserstoff-Forschung Baden-Württemberg (ZSW), Stuttgart, Germany In cooperation with: - 0 -
Outline Machine Learning (ML) for RE applications The German-Chilean project PROREC Wind power forecasting with ML Solar power forecasting with ML Integration of RE into the energy system - 1 -
What is Machine Learning (ML) good for? ML methods excel at fitting parameters from data. But even if the problem is not understood well enough to solve physically, a heuristic model can be built, provided we have data. physical or engineering model: known structure controlled experiments generic model (NN, regression, histogram, ): little prior knowledge exploratory data analysis - 2 -
- 3 - Why isn t everyone using ML?
Outline Machine Learning (ML) for RE applications The German-Chilean project PROREC Wind power forecasting with ML Solar power forecasting with ML Integration of RE into the energy system - 4 -
Chile activities: Status and goals Phase 1: WindSageCL (2012 2014) getting to know the country first contacts and wind forecasts CL Phase 2: PROREC (2015 2018) develop PV- and improve wind forecasts simulation of NCRE supply Phase 3:??? hydrogen from NCRE extension to other S-American countries - 5 -
PROREC project structure Prediction of and Optimized Provision with Renewable Energy in Chile Oct. 2015 to Aug. 2018-6 -
Outline Machine Learning (ML) for RE applications The German-Chilean project PROREC Wind power forecasting with ML Solar power forecasting with ML Integration of RE into the energy system - 7 -
Forecast model setup input vector inputs + targets = 1 pattern split available patterns into training set and independent(!) test set All results shown stem from the test set! possibly other weather models wind forecast targets Deep Neural Network (DNN) Training and test data sets - 8 -
Training on GPU hardware 4 Titan + 8 Titan Black 8 TitanX + 8 TitanXP 1x K20, some older models ca. 220 TFLOPS - 9 -
Example results of Wind Power Forecast in Chile Monte Redondo wind farm observación pronóstico Punta Colorada wind farm
Totoral wind farm Input data sensitivity
Probabilistic power forecast Different target encoding => DNN learns variability from training dataset
Probabilistic forecasts as percentile plots Conversion to percentiles makes probabilistic forecast easier to interpret => adds useful information for traders, grid operators, Totoral wind farm 2015-07-01 12:00 UTC e.g. 90% likelihood that power lies in this band 2015-11-21 16:00 UTC
Outline Machine Learning (ML) for RE applications The German-Chilean project PROREC Wind power forecasting with ML Solar power forecasting with ML Integration of RE into the energy system - 14 -
Cascading forecast time scales escala de tiempo 1 s 1 min 1 h 1 d 1 mes cloud camera satellite images numerical weather prediction work in progress - 15 -
Finding Best Network Architecture Maria Elena PV farm Slight preference for 2 hidden layers with 64 neurons each 2015 2016 2 yrs of hourly data from CEN, manual QC
Solar Forecast Percentiles Luna PV farm observation forecast 120h! Event missed in deterministic forecast, but percentiles show significant drop!
Solar Forecast Percentiles Luna PV farm observation forecast Deterministic forecast would look the same below and above!
Error statistics for different PV farms Input: Three global NWP models, plus live data => Hi-Res NWP model would probably improve results! - 19 -
Cascading forecast time scales escala de tiempo 1 s 1 min 1 h 1 d 1 mes cloud camera satellite images numerical weather prediction work in progress - 20 -
Preparing the image data Inverse polar transformation, because horizon more important than zenith NB: This is not the same image!
Neural network model setup
Neural network model setup Source: www.mdpi.com/1099-4300/19/6/242 Source: neuralnetworksanddeeplearning.com
Presenting the image data concatenated sequential stacked
Solar Power Forecasting using Cloud Camera (5 min forecast of PV power @ Widderstall, Germany) variability of the cloud situation - 28 -
- 29 - Adding satellite images
Compression of Satellite Images with Deep Autoencoder Needed to combine satellites with other sources (data volume!) Special DNN constructed, with 14(!) layers Achieved compression ratio of 32 (using ZCodes of the autoencoder) z - 30 -
- 31 - Cascaded forecast model
- 32 - Combination model for short-term forecasts
Outline Machine Learning (ML) for RE applications The German-Chilean project PROREC Wind power forecasting with ML Solar power forecasting with ML Integration of RE into the energy system - 33 -
- 34 - P²IONEER (Hybrid Power Plant Simulation)
What is Machine Learning (ML) good for? ML methods excel at fitting parameters from data. But even if the problem is not understood well enough to solve physically, a heuristic model can be built, provided we have data. physical or engineering model: known structure controlled experiments generic model (NN, regression, histogram, ): little prior knowledge exploratory data analysis - 35 -
Comparison of assumptions CapEx 2016 (US$/kW) CapEx 2018 (US$/kW) OpEx (invest/a) Wind 2,200 1,800 3.5 % Solar PV 1,600 850 1.5 % Li-Ion 650 550 1.0 % CHP 980 980 2.8 % Diesel 600 600 2.2 % Demand profiles from CEN and GIZ Supply profiles modeled according to Explorador Eólica y Solar Average market prices: Diesel 2016: 5.9 ct/kwh 2018: 4.5 ct/kwh Electricity 2016: 14.3 ct/kwh 2018: 5.67 ct/kwh sources: NRDC, SOWITEC, ZSW
power [MW] Simulation of Hybrid Energy Systems Example Demand has to be met at all hours of the year Excess energy goes into storage and/or is traded over the grid - 37 -
Simulation results for some mines Overall large reduction compared to 2016 run. Cheaper solar power can not quite compensate drop in fuel and electricity prices. Optimization criteria may not be LCOE of electricity alone. => depends on problem! - 38 -
P²IONEER optimization result Here we optimized the supply for a region in Germany, including a local heat distribution network. - 39 -
P 2 IONEER with Power-to-Water component
Conclusions Excellent cooperation between ZSW and ULS produced many interesting results. ML is a powerful tool for dealing with RE data, if applied properly. Forecast quality for wind an PV power depends on availability and quality of local measurements and NWP models. Adding cloud cameras and/or satellite data to improve PV forecasts is non-trivial but probably worthwhile. Interplay of different RE and conventional energy sources, storage and sinks can be simulated with P²IONEER, but results again depend on reliable input data. - 41 -
// Energy with a future // Zentrum für Sonnenenergie- und Wasserstoff- Forschung Baden-Württemberg (ZSW) Thank you for your attention! Contact: martin.felder@zsw-bw.de Stuttgart: Photovoltaics Division (with Solab), Energy Policy and Energy Carriers, Central Division Finance, Human Resources and Legal Widderstall: Solar Test Facility Ulm: Electrochemical Energy Technologies Division, Main Building & elab - 42 - - 42 -