Predicting Solar Generation From Weather Forecasts Using .
Predicting Solar Generation from WeatherForecasts Using Machine LearningNavin Sharma, Pranshu Sharma, David Irwin, and Prashant ShenoyDepartment of Computer ScienceUniversity of Massachusetts AmherstAmherst, Massachusetts Abstract—A key goal of smart grid initiatives is significantlyincreasing the fraction of grid energy contributed by renewables.One challenge with integrating renewables into the grid is thattheir power generation is intermittent and uncontrollable. Thus,predicting future renewable generation is important, since thegrid must dispatch generators to satisfy demand as generationvaries. While manually developing sophisticated prediction models may be feasible for large-scale solar farms, developing themfor distributed generation at millions of homes throughout thegrid is a challenging problem. To address the problem, in thispaper, we explore automatically creating site-specific predictionmodels for solar power generation from National Weather Service(NWS) weather forecasts using machine learning techniques. Wecompare multiple regression techniques for generating predictionmodels, including linear least squares and support vector machines using multiple kernel functions. We evaluate the accuracyof each model using historical NWS forecasts and solar intensityreadings from a weather station deployment for nearly a year.Our results show that SVM-based prediction models built usingseven distinct weather forecast metrics are 27% more accuratefor our site than existing forecast-based models.I. I NTRODUCTIONA key goal of smart grid efforts is to substantially increasethe penetration of environmentally-friendly renewable energysources, such as solar and wind. For example, the RenewablesPortfolio Standard targets up to 25% of energy generationfrom intermittent renewables [1], while Executive Order S21-09 in California calls for 33% of their generation to comefrom renewables by 2020 [2]. Substantial grid integrationof renewables is challenging, since their power generationis intermittent and uncontrollable. The modern electric gridpermits households to consume electricity in essentially arbitrary quantities at any time, and is not currently designed forvast quantities of uncontrollable generation. Instead, the gridconstantly monitors the demand for electricity, and dispatchesgenerators to satisfy demand as it rises and falls. Fortunately,electricity demand is highly predictable when aggregating overthousands of buildings and homes. As a result, today’s gridis able to accurately plan in advance which generators todispatch, and when, to satisfy demand.The problem with substantial renewable integration is thatthe electricity renewables generate is not easily predictablein advance and varies based on both weather conditions andsite-specific conditions. While utilities may take the timeto manually develop accurate prediction models for large-scale solar farms that produce multiple megawatts, manuallydeveloping specialized models that predict the power outputfrom distributed generation at many small-scale facilities atsmart homes and buildings throughout the grid is infeasible.This fact is evident in current net metering laws for most states,which allow consumers to sell energy produced from on-siterenewables back to the grid, but typically places low caps onboth the total number of participating customers and/or thetotal amount of energy contributed per customer [3]. As oneexample, Massachusetts caps the total number of participatingcustomers at 1% of all customers. Utilities restrict the contribution from renewables, since, unlike electricity demand,renewable generation is not easily predictable, and complicatesadvance planning of the grid’s generator dispatch schedule.To facilitate better planning and lower the barrier to increasing the fraction of renewables in the grid, we focus on theproblem of automatically generating models that accuratelypredict renewable generation using National Weather Service(NWS) weather forecasts. Specifically, we experiment with avariety of machine learning techniques to develop predictionmodels using historical NWS forecast data, and correlatethem with generation data from solar panels. Once trainedon historical forecast and generation data, our predictionmodels use NWS forecasts for a small region to predict futuregeneration over several time horizons. Our experiments inthis paper use solar intensity as a proxy for solar generation, since it is proportional to solar power harvesting [4].Importantly, since we generate our models from historical sitespecific observational power generation data, they inherentlyincorporate the effects of local characteristics on each site’scapability to generate power, such as shade from surroundingtrees. Since local characteristics influence power generation,individual sites must tune prediction models for site-specificcharacteristics. We view automatic model generation as criticalto scaling distributed generation from renewables to millionsof homes throughout the grid.Our goal is to automate generating prediction models forsmart homes that include on-site renewables. Both the grid andindividual smart homes may use these prediction models foradvance planning of electricity generation and consumption.The grid can use the models to plan generator dispatchschedules in advance as the fraction of renewables increases inthe grid. Smart homes can use the models to potentially plan
their consumption patterns to better match the power that theygenerate on-site. In both cases, better prediction models are aprerequisite for increasing efficiency and encouraging broaderadoption of distributed generation from renewables in the gridand at smart homes. In studying prediction models for solarenergy harvesting, we make the following contributions. Data Analysis. We analyze extensive traces of historicaldata from a weather station, as well as the correspondingNWS weather forecasts, to correlate the weather metricspresent in the forecast with the solar intensity, in wattsper m2 , recorded by the weather station. Our analysisquantifies how each forecast parameter affects each otherand the solar intensity. For solar energy harvesting, wefind that sky cover, relative humidity, and precipitationare highly correlated with each other and with solarintensity, while temperature, dew point, and wind speedare only partially correlated with each other and withsolar intensity. Model Generation. We apply multiple machine learningtechniques to derive prediction models for solar intensityusing multiple forecast metrics, and then analyze theprediction accuracy of each model. We use machinelearning on a training data set of historical solar intensityobservations and forecasts to derive a function that computes future solar intensity for a given time horizon froma set of forecasted weather metrics. We formulate modelsbased on linear least squares regression, as well as supportvector machines (SVM). We find that SVM with radialbasis function kernels built using historical data fromseven weather metrics is 27% more accurate than existingforecast-based models that use only sky condition forpredictions [4] and is 51% better than simple approachesthat only use the past to predict the future.In Section 2 we analyze forecast metrics and explore howthey affect each other, as well as how they affect solarintensity, while in Section 4 we describe and evaluate multiplemachine learning strategies for generating prediction modelsusing our weather station data and NWS forecasts. Finally,Section 5 discusses related work and Section 6 concludes.II. DATA A NALYSISWe collect weather forecast data and observational solar intensity data for 10 months starting from January2010. We obtain historical forecast data from the NWS athttp://www.weather.gov, which we have been collecting forthe past 2 years. The NWS provides historical textual forecastsfor small city-size regions throughout the U.S., which includemultiple weather metrics for every hour of every day for thelast few years. Each forecast includes predictions of eachmetric every 1 hour from 1 hour to 6 days into the future.Examples of weather metrics include temperature, dew point,wind speed, sky cover, probability of precipitation, and relativehumidity. Sky cover is an estimate of the percentage (0%100%) of cloud coverage in the atmosphere. In addition tomaking historical forecasts available, the NWS also operatesa real-time web service that enables applications to retrieveSunny WinterDayCloudy SummerDayFig. 1. Solar intensity shows seasonal variation with days of a year, althoughdaily weather conditions also have a significant impact.forecasts programmatically as they become available. In addition to these metrics, we include the specific day of the yearand time of the day as metrics, since daylight influences solarintensity and varies throughout the year for a given location.We use observational solar intensity data from an extendedweather station deployment on the roof of the Computer Science Building at the University of Massachusetts Amherst. Theweather station reports solar intensity in watts/m2 every 5 minutes of every day. Traces from the weather station are availableat http://traces.cs.umass.edu. As we show in previous work,power generation from solar panels is directly proportional tosolar intensity [4]; in general, solar panel inefficiencies resultin power output that is a fixed percentage decrease from theraw solar intensity readings at the same location. We use NWSforecasts for Amherst, Massachusetts. In this section, we studyhow solar intensity varies with individual forecast parametersand how these forecast parameters are related to each other.The purpose of our data analysis is to provide intuition intohow solar intensity and solar panel power generation dependson a combination of multiple weather metrics, and is not easilypredictable from a single weather metric. The complexity inpredicting solar intensity from one or more weather metricsmotivates our study of automatically generating predictionmodels using machine learning techniques in the next section.Fig. 1 shows how the day of the year affects solar intensityby charting the average solar intensity reading at noon perday over our 10 month monitoring period, where day zerois January 1st, 2010. As expected, the graph shows that thesolar intensity is lowest in January near the winter solstice andincreases into the summer before decreasing after the vernalequinox. Additionally, the graph also implies that other conditions also have a significant impact on solar intensity, sincemany days throughout the spring and summer have low solarintensity readings. The graph shows that solar intensity and theday of the year are roughly correlated: most of the time, butnot always, a summer day will have a higher solar intensitythan a winter day. However, other factors must contribute tothe solar intensity, since there are clearly some sunny winterdays that record higher solar intensity readings than some
(a)(b)(c)Fig. 2. Solar intensity and wind speed show little correlation (a). Solar intensity shows some correlation with temperature at high temperatures (b) and withdew point at high dew points (c).(a)Fig. 3.(b)(c)Solar intensity generally decreases with increasing values of sky cover (a), relative humidity (b), and precipitation potential (c).cloudy summer days. To better understand correlations withother weather metrics, we model similar relationships for theother forecast metrics.For example, Fig. 2 shows that wind speed, dew point,and temperature are not highly correlated with solar intensity.Solar intensity varies almost uniformly from lower to highervalues at any value of wind speed (a). Thus, wind speedhas nearly zero correlation with solar intensity and its valueis not indicative of the solar intensity or solar panel powergeneration. Both temperature (b) and dew point (c) correlatewith solar intensity at higher values: if the temperature or dewpoint is high, then the solar intensity is likely to be high.However, if the temperature or dew point is low, the solarintensity exhibits a more significant variation between highand low values. The results are intuitive. For example, in thesummer a high temperature is often dependent on sunlight,while in the winter sunlight contributes less in raising theambient temperature.In contrast, Fig. 3 shows that sky cover, relative humidity,and chance of precipitation have high negative correlationswith solar intensity. In each case, as the value of the metric increases, the solar intensity reading generally decreases.However, as with the day of the year, there must be otherfactors that contribute to the solar intensity reading, since thereare some days with a high sky cover, relative humidity, andprecipitation probability, but a high solar intensity reading andvice versa. In addition to exhibiting complex relationships withsolar intensity, each weather metric also exhibits a complexrelationship with other weather metrics. For example, Fig. 4shows that relative humidity (a) and chance of precipitation(b) exhibit strong, but not perfect correlations, with sky cover,while relative humidity is strongly correlated with chance ofprecipitation (c). In all three cases, the metrics rise in tandem,although the relationship is noisy due to the value of otherweather metrics.Table 1 shows correlation coefficients for each weathermetric using the Pearson product-moment correlation coefficient, which divides the covariance of the two variablesby the product of their standard deviations. The higher theabsolute value of the correlation coefficient, the strongerthe correlation between the two weather metrics—a positivecorrelation indicates an increasing linear relationship, while anegative correlation indicates a decreasing linear relationship.The complex relationships between weather metrics and solarintensity shown in this table motivate our study of automatedprediction models using machine learning techniques in thenext section.III. P REDICTION M ODELSWe represent both observational and forecast weather metrics as a time-series that changes due to changing weatherpatterns and seasons. As the previous section shows, solarintensity depends on multiple weather metrics, which complicates the task of developing an accurate prediction model. Thehigh dimensionality of the time-series data motivates our studyof regression methods to develop solar intensity predictionmodels. To generate each model we provide eight monthsof training data (January to August) as input, which includessolar intensity readings as well as NWS forecasts for 6 weathermetrics. The machine learning techniques automatically output
(a)(b)(c)Fig. 4. Relative humidity (a) and precipitation % (b) positively correlate with sky cover. Relative humidity also increases with increasing precipitation % (c).TABLE IC ORRELATION MATRIX SHOWING CORRELATION BETWEEN DIFFERENT FORECAST PARAMETERS .DayTemp.Dew pointWindSky 00-0.01920.0340-0.1025Sky 0a function that computes solar intensity from the 6 weathermetrics, as well as the day of the year. We use the remaining2 months of our data set to test the model’s accuracy. Onebenefit of using machine learning to automatically generateprediction models is that, in general, the more training datathat is available, the more accurate the model.We focus our study on short-term forecasts three hoursin the future. For our experiments, we develop models thatdetermine a relationship at any time t between the solarintensity and the forecast weather metrics three hours in thepast (t 3). Note that we are able to apply our techniques toforecasts of any length; we choose three hours as a simpleillustration. Using our models and the three hour forecast,we are able to compute a prediction for the solar intensitythree hours in the future. The models we generate are simplefunctions, of the form below, that compute solar intensity frommultiple weather metrics including the day of the year. Wecould also add time of the day as an additional metric, butfor ease of exposition our experiments focus on predictionsat noon. We compare the accuracy of our models with eachother, as well as against a simple model we developed inprior work [4] based solely on the sky condition metricand against a simple past-predicts-future model. Our previousmodel multiplies the maximum power a solar panel is able togenerate at a given time (of the day and year) by (1-SkyCover),since sky cover represents an estimate of the percentage of theatmosphere the sun is covering.in degrees Fahrenheit, wind speed in miles per hour, sky coverin percentage between 0% and 100%, precipitation potential inpercentage between 0% and 100%, and humidity in percentagebetween 0% and 100%. However, before applying any regression techniques below we normalize all feature data to havezero mean and unit variance. To quantify the accuracy of eachmodel, we use the Root Mean Squared Error (RMS-Error)between our predicted solar intensity at any time and the actualsolar intensity observed. RMS-Error is a well-known statisticalmeasure of the accuracy of values predicted by a time-seriesmodel with respect to the observed values. An RMS-Error ofzero indicates that the model exactly predicts solar intensitythree hours in the future. The closer the RMS-Error is to zerothe more accurate the model’s predictions.SolarIntensity F(Day, Temperature, DewPoint, WindSpeed,SkyCover, Precipitation, Humidity)SolarIntensity 1.18*Day 77.9*Temp 33.11*DewPoint 22.8*WindSpeed - 96.9*SkyCover - 49.15*Precipitation 43.4*HumidityF is the function that we determine using different regression methods. We preserve the units of each metric: werepresent each day as a value between 0 and 365, temperatureA. Linear Least Squares RegressionWe first apply a linear least squares regression method topredict solar intensity. Linear least squares regression is a simple and commonly-used technique to estimate the relationshipbetween a dependent or response variable, e.g., solar intensity,and a set of independent variables or predictors. The regressionminimizes the sum of the squared differences between theobserved solar intensity and the solar intensity predicted by alinear approximation of the forecast weather metrics. Applyingthe linear least squares method to the eight months of trainingdata yields the prediction model below, with coefficients foreach metric.We verify the prediction accuracy using our test datasetfor the remaining months of the year. We observe the cross
1000ObservedLinear regression800Solar Intensity (watts/m2)Solar Intensity (watts/m 040Days (starting 09/01/2010)5001020304050Days (starting 09/01/2010)Fig. 5. Observed and predicted solar intensity using linear least squaresregression for September and October 2010.Fig. 6. Observed and predicted solar intensity, using SVM regression withan RBF kernel, for the months of September and October 2010.validation RMS-Error and prediction RMS-Error in the solarintensity as 165 watts/m2 and 130 watts/m2 , respectively. Wecross validate the regression model with the training dataset(from Jan. to Aug. 2010) and verify its prediction accuracyusing the testing dataset (Sept. and Oct. 2010). The crossvalidation RMS-Error quantifies how well the model predictsvalues in the training data set, while the prediction RMS-Errorpredicts how well the model predicts values in the testing dataset. Fig. 5 shows the observed and predicted solar intensity forSeptember and Octobe
Predicting Solar Generation from Weather Forecasts Using Machine Learning Navin Sharma, Pranshu Sharma, David Irwin, and Prashant Shenoy Department of Computer Science University of Massachusetts Amherst Amherst, Massachusetts 01003 {nksharma,pranshus,irwin,shenoy}@cs.umass.edu Abstract—A key goal of smart grid initiatives is significantly
Weather instruments are used to measure and record the weather. Weather instruments can be found in weather stations on land. The Met Office has hundreds of weather stations all over the UK. Weather instruments are also found at sea. They are found on some ships, but mainly on weather buoys designed to monitor weather and sea conditions.
Solar Milellennium, Solar I 500 I CEC/BLM LLC Trough 3 I Ridgecrest Solar Power Project BLM 250 CEC/BLM 'C·' ' Solar 250 CEO NextEra I Trough -----Abengoa Solar, Inc. I Solar I 250 I CEC Trough -I, II, IV, VIII BLM lvanpah SEGS Solar I 400 I CECJBLM Towe'r ico Solar (Solar 1) BLM Solar I
Mohave/Harper Lake Solar Abengoa Solar Inc, LADWP San Bernardino County 250 MW Solar Trough Project Genesis NextEra Energy Riverside County 250 MW Solar Trough Beacon Solar Energy Project Beacon Solar LLC Kern County 250 MW Solar Trough Solar Millennium Ridgecrest Solar Millenn
of solar generation (via solar intensity) using simple, readily-available weather data. By limiting the uncertainty of predicted solar forecasts, such a model has the potential to allow grid sites to reduce
Deep Learning Technology for Predicting Solar Flares from (Geostationary Operational Environmental Satellite) Data . National Space Weather Program as “Conditions on the Sun and in the solar wind, magnetosphere, ionosphere and . rate of generation of data monitoring solar activity, scientists
4. Solar panel energy rating (i.e. wattage, voltage and amperage). DESIGN OF SYSTEM COMPONENTS Solar Panels 1. Solar Insolation Solar panels receive solar radiation. Solar insolation is the measure of the amount of solar radiation received and is recorded in units of kilowatt-hours per square meter per day (kWh/m2/day). Solar insolation varies .
responding to the solar direction. The solar tracker can be used for several application such as solar cells, solar day-lighting system and solar thermal arrays. The solar tracker is very useful for device that needs more sunlight for higher efficiency such as solar cell. Many of the solar panels had been
hubungan antara asupan asam folat dengan kadar Hb dengan nilai p 0,64. Kata Kunci : asupan fe, asupan folat, kadar hb, tb paru . Abstract . Tuberculosis pulmonary can lead to various metabolic disorders and system disturbances in the body, one of which is synthetic disorder of Hemoglobin levels. Some nutrients which can influence the synthetic of Hemoglobin levels are iron (Fe) and folic .
