Probabilistic And Point Solar Forecasting Using Attention .

Probabilistic and Point Solar Forecasting UsingAttention-Based Dilated Convolutional Neural NetworkMoumita Saha, Bhalchandra Naik, Claire MonteleoniDepartment of Computer ScienceUniversity of Colorado Boulder, USAEGU General Assembly, May 2020Sharing Geoscience Online

Scope1Introduction2CNN Based Solar Prediction Model3Experimental Results4Conclusion & Future Work

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkSolar IrradianceSolar is a good source for renewable and clean energySolar Irradiance is the flux of radiant energy received per unitarea of the earthSolar irradiance has many significantapplications:the prediction of energy generationfrom solar power plantthe heating and cooling loads ofbuildingsclimate modeling and weatherforecasting

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkImportance of Solar IrradianceClimate change is evident and netdamages are quite significantA high rise in greenhouse gases is amajor cause for climate changeContributed with burning of fossil fuels and otheranthropogenic activitiesRenewable energy sources like solar are a good source forclean energy production for combating climate change

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkForecasting Solar IrradianceEfficient integration of solar energy intoelectrical grids requires an accurateprediction of solar irradianceInfluences the production of solar energyat photo-voltaic plantAccurate prediction of irradiance helps in forecasting theenergy production at a lead timeNeed for crucial decisions in scheduling harvesting andestimating power requirements for the future

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkConvolutional Neural Network for Solar IrradianceWe aim to forecast downwelling global solar irradiance using afixed history of the variablesWe propose a convolutional neural network with dilated kerneland attention-based mechanism for predicting the solarirradianceWe present both probabilistic and point forecasts of solarirradiance at multiple lead timesForecast for all four seasons: Fall, Winter, Spring, andSummer

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkConvolutional Neural NetworkConvolutional neural networks (CNN) are capable ofextracting features from data that have local spatial relationsWe use CNN to map samples of sub-sequences from atime-series series to some observed value in the future

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkDilated Kernel and Temporal Attention MethodDilation refers to cavities in the kernel,which allows looking for dependencies innon-adjacent cellsWe added dilation for capturinglong-term dependenciesAttention mechanism compels the model to focus on the partsof the input that bear a high impact on the outputIn our case, it emphasis the past states of input which havethe highest impacts on the output solar irradiance

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkModel Training and Loss FunctionThe input dimension is twenty-one and the output dimensionis either the number of quantiles (probabilistic) or one (point)The quantile loss QLα used in probabilistic forecasting is((y ŷ )(1 α), if (y ŷ ) 0QLα (ŷ , y q) (y ŷ )α,if (y ŷ ) 0For point-forecasting, we use a L2 loss functionPL(y , ŷ ) (y ŷ )2

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkPerformance MetricsRoot mean square error (RMSE) for point forecastingsPT2t 1 (yt yˆt )RMSE TContinuous Ranked Probability Score (CRPS) for probabilisticforecastingZCRPS 01T1 XQLα (yt , yˆt )dαTt 1

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkBaseline Persistence ModelsSimple persistence (SP) model can be defined as:Ip (t 4t) I (t),where I (t) is the solar irradiance at current timeSmart persistence (SMP) model forecast the irradiance bymultiplying the clear-sky index by the future clear-skyirradianceIsp (t 4t) kt (t) Iclr (t 4t)where kt (t) is the clear-sky index at current time and Iclrdenotes the clear-sky irradiance

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkPoint Forecasting for Solar IrradianceThe model shows higher performance for the fall and winterseasonsRMSE for point solar irradiance forecasting by CNN and simple persistence (SP)models at two different leadsLeads3 hrs6 hrsFallCNNSP1693251833753 hrs6 NSP122280234369100263267464Fort 375238491202252326383

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkPoint Forecasting for Solar IrradianceSolar Irradiance (Watts/ m 2)Hourly observed and predicted solar irradiance for Boulder winter seasonObserved solar irradiancePredicted solar irradiance4003002001000102030Hourly period4050

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkProbabilistic Forecasting of Solar IrradianceWe use quantile regression for probabilistic forecastingWe predicted solar at 19 different quantiles withα [0.05, 0.10, 0.15, ., 0.95]Similar high performance observed for the winter and fallCRPS for probabilistic solar irradiance forecastingLeads3 hrsFall182.33 hrs138.3Boulder-ColoradoWinterSpring130.4242.2Fort Peck-Montana122.3181.6Summer261.7235.8

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkSolar Forecasting by CNN and Persistence Models

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkSkill Score of CNN Model over Persistence ModelsCNN model shows high skill score over the baselinepersistence models for both the lead timesSkill score for point solar irradiance forecasting over persistence modelsBoulder-ColoradoSimple persistenceSmart persistenceLeadsFallWinter Spring Summer FallWinter Spring Summer3 hrs47.956.336.5358.511.8-3.66.26 hrs51.861.842.451.513.316.99.511.6Fort Peck-Montana3 hrs45.336.144.824.28.211.76.47.26 hrs47.123.950.038.713.215.29.211.6

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkConclusionThe CNN model learns the mapping from the past time-seriesof climatic variables to solar irradiance and predicts theirradiance at multiple lead timesDilated kernel and temporal attention aid to boost theforecast accuracyAdding past irradiance as an input improves the forecastSolar irradiance prediction for winter and fall seasons arebetter than other seasons

IntroductionCNN Based Solar Prediction ModelExperimental ResultsConclusion & Future WorkFuture WorkApply a dual-stage attention mechanism which can learn theimportance of input features and temporal historyAn adaptive temporal attention method can aid in detectingthe number of attention steps to be consideredSolar irradiance in finer grids over the United States tosupport power plants in crucial decision-making

References1‘Validation of short and medium term operational solar radiation forecasts in theUS,’Solar Energy, 20102‘Short-term probabilistic forecasts for direct normal irradiance,’ RenewableEnergy, 20173‘Online short-term solar power forecasting,’ Solar Energy, 20094‘The SURFRAD Network. Monitoring Surface Radiation in the ContinentalUnited States,’ 20105‘Day-ahead solar irradiance forecasting for microgrids using a long short-termmemory recurrent neural network:A deep learning approach,’ Energies, 20196‘CSRNet: Dilated convolutional neural networks for understanding the highlycongested scenes,’ CVPR, 20187‘Neural machine translation by jointly learning to align and translate,’arXiv:1409.0473, 20148‘A dual-stage attention-based recurrent neural network for time seriesprediction,’ IJCAI, 2017

.Thank you

Solar is a good source for renewable and clean energy Solar Irradiance is the ux of radiant energy received per unit area of the earth Solar irradiance has many signi cant applications: the prediction of energy generation from solar power plant the heating and cooling loads of buildings climate modeling and weather forecasting