Weather Based Photovoltaic Energy Generation Prediction .

Weather Based Photovoltaic Energy GenerationPrediction Using LSTM NetworksSahar ArshiFaculty of Engineering and EnvironmentUniversity of Northumbria,Newcastle, UKsahar.arshi@northumbria.ac.ukLi ZhangFaculty of Engineering and EnvironmentUniversity of NorthumbriaNewcastle, UKli.zhang@northumbria.ac.ukRebecca StrachanFaculty of Engineering and EnvironmentUniversity of NorthumbriaNewcastle, UKrebecca.strachan@northumbria.ac.ukAbstract— Photovoltaic (PV) systems use the sunlight andconvert it to electrical power. It is predicted that by 2023, 371,000PV installations will be embedded in power networks in the UK.This may increase the risk of voltage rise which has adverseimpacts on the power network. The balance maintenance isimportant for high security of the physical electrical systems andthe operation economy. Therefore, the prediction of the output ofPV systems is of great importance. The output of a PV systemhighly depends on local environmental conditions. These includesun radiation, temperature, and humidity. In this research, theimportance of various weather factors are studied. The weatherattributes are subsequently employed for the prediction of thesolar panel power generation from a time-series database. LongShort Term Memory networks are employed for obtaining thedependencies between various elements of the weather conditionsand the PV energy metrics. Evaluation results indicate theefficiency of the deep networks for energy generation prediction.A feed-in-tariff scheme was proposed by UK in 2010 withthe main purpose of reducing UK’s carbon emission [1]. Inaccordance to the feed-in-tariff scheme, a low carbon networkproject was defined in 2015, entitled “Validation ofPhotovoltaic Connection Assessment Tool”. The projectinvolved collecting real-life data for trial purposes, as well asensuring that the connection assessment tools were fit forpurpose [1]. There were some other objectives and benefitsrelated to the project at the time. These include understandingthe underlying associated PV generator behaviours by dataanalysis. The other target was to identify the underlying issuesin the connection procedures. The project was successful inachieving a validated and practical connection assessmentapproach. As a cross product, a rich dataset was producedwhich is useful for distribution network operators andacademic institutions. The produced database is deposited inLondon datastore and provides the basis for this research study.Keywords—Photovoltaic systems, Solar panels, Long ShortTerm Memory, Energy Forecasting.In this research, the PV energy generation database will bestudied from a different viewpoint rather that the initial targetsof the UK Networks project (as mentioned above). Theresearch is in accordance with the related native weatherconditions. The output of a PV system highly depends on localenvironmental circumstances as well as the sun radiation. Thesunlight can be occluded by the clouds. The wind blowingdirection can bring clouds and effect the way the panel receivesthe sunlight. There is clearly a correlation between solarradiation and PV outputs. It is often assumed that a PV systemis likely to have its maximum output on a clear sunny day.Nevertheless, the temperature is also an important factor on theefficiency of PV generators. The work in [1] demonstrated thatPV generators could potentially produce higher levels ofelectricity on cloudy days. These kinds of analysis results mayoften be in contradictory with the common beliefs. Therefore,different environmental factors impact the overall energygeneration divergently.I. INTRODUCTIONThe usage of solar power generators has been encouragedin recent years owing to various environmental benefits. Theforecasting of Photovoltaic (PV) power is thus an uprisingresearch branch. However, the power generated by the PVsolar devices are highly affected by weather conditions. Theenergy generated by photovoltaic systems is variable in termsof seasonality and weather factors. The solar panel orientation(East facing or West facing) affects PV output. The imported,exported, and generated energy are important terminologiesused in the field. The imported energy refers to the electricitytaken from the network by the household, whereas the exportedenergy refers to the electricity generated, but not consumed bythe household [1]. The energy exported is the excess electricitywhich is injected to the network. The prediction of thegenerated energy is of great importance to grid operatorswhose responsibility is to keep the imported and exportedenergy balanced in the distribution system [2]–[5].XXX-X-XXXX-XXXX-X/XX/ XX.00 20XX IEEEIn this paper, various weather related data are analysed. Anattribute selection procedure is employed for identifying themost important weather factors which affect the generatedpower. Later, a Long-Short Term Memory (LSTM) network

model is employed for predicting the PV systems poweroutput. The LSTM architecture is equipped with memory unitsthat can be useful in forecasting the temporal variations of PVgenerated power metrics.II. LITERATURE REVIEWA. Related Research for Predicting Solar PowerGenerationPrediction of PV generated power can be traced back invarious studies. The previous approaches in the area of powergeneration forecasting can be categorized into four majorgroups including AI, physical, statistical, and hybrid methods.AI models are one of the popular methods for forecastingthe outputs of PV plants. Various machine learning modelshave been used ever since. Some of the previous research inthe area applied artificial neural networks, while recent studiestarget tools such as deep learning models. Some examples ofAI applications in solar radiation forecasting for PV outputprediction can be found in [4], [6], [7]. We discuss some of thestudies in detail below.At the University of Illinois, the researchers applied somelinear and non-linear machine learning algorithms forforecasting the solar energy generation [8]. They recorded thesolar panels outputs at the campus of Illinois University. Theirwork used weather information and selected day lightobservations. The employed methods included weighted linearregression trees and LSTM. The later achieved the bestperformance in their application. The study also found thetime-series correlations between all the weather attributes andactual energy output. They found that the cloud coverage,humidity, visibility and dewpoint were the most importantfeatures for solar panels output forecast. A similar research in[8] also stated that they could have taken advantage of weatherdependency in a way that previous weather conditions wouldaffect the current weather conditions in their model. Instead,they randomized the training dataset.In [6] , Meyers applied the 3 hourly and 5 minutes timestep intervals to identify the ramp events or deviation from along-term trend within a short-term period. Their work usedauto-regressive model, K-nearest neighbour and artificialneural networks.In [9], the researchers suggested the usage of LSTMRecurrent Neural Network (LSTM-RNN) to predict thegenerated power of PV systems using a time-series databasewith one varying feature. They employed the hourly intervaldataset over a period of a year’s time. The LSTM modeloutperformed other methods using the one-featured time-series.The work used various LSTM architectures as well as basicLSTM for regression by using window techniques (LSTM withmemory between batches). Their database comes from twosites in Egypt. Other machine learning tools were also appliedin their study. However, the LSTM method achieved the bestprediction in terms of lower mean absolute error and meansquare error compared to those of other methods they applied.Besides machine learning methods, physical models werealso applied for predicting PV system output. The usage ofsatellites and numerical weather prediction models are two ofthe physical models which were used in the past for forecastingthe solar radiation and indirectly predicting the output of PVsystems. The research in the solar power prediction isintertwined with weather forecasting approaches, as the lateraffects the outcome of the energy generated by PV systems. In[10], the numerical weather prediction models were employedfor photovoltaic and solar power generation forecasting. Inmanand colleagues in [11] provided a comprehensive literaturereview on solar forecasting methods for renewable energyintegration. They discussed the prediction method for solarresources (weather conditions forecasting) and the poweroutput of the solar plants. Some other recent review works onsolar power forecasting can be found in [12].Moreover, statistical methods have also been used forpredicting PV outputs. Statistical models are usually based onthe short-term historical data. Auto-Regressive-IntegratedMoving-Average models (ARIMA) are examples of statisticalmodels [13]–[15]. Forecasting the PV output time-series datausing ARIMA models was among the first collections ofresearch in this area. In [16], ARMA model with differentparametrization are employed to analyse and forecast theresiduals in daily solar radiation time-series in Malaysia sites.Finally, hybrid methods combine various methods togetherfor prediction. For instance, AI models were integrated withphysical methods for PV system output prediction. The studyin [17] discussed on how the sky images were analysed forforecasting the output of Photovoltaic facility plant in Nevadausing image processing techniques.In this research, an LSTM model as a deep learning modelwas employed for modelling the time-series dependenciesbetween weather condition changes and solar power generationmetrics. The predictions are performed up to 48 time-steps(number of unrollings). Moreover, more than one solar powerfactors were predicted in a multistep-ahead predictionapproach. The model also embeded the dependency pertainingto weather conditions changes.B. Long Short Term Memory NetworksThe LSTM architecture was originally proposed by SeppHochreiter and Jurgen Schmidhuber in 1997 [18]. An LSTMnetwork is a Recurrent Neural Network (RNN) [19] whichconsists of memory units. An LSTM network refines thevanishing gradient problem associated with RNNs [20]. AnLSTM cell consists of a memory cell and multiplicative gateswhich work as regulators. LSTM networks learn short-term orlong-term dependencies between elements of time-series data[5]. This characteristic makes LSTM networks suitable formaking predictions.An LSTM cell is similar to RNN in its recurrentarchitecture. However, LSTM cells are equipped with amemory cell which makes them more desirable for broaderranges of applications. The LSTM architecture includes somegates as well as the input gate (𝐼𝐼𝑡𝑡 ), output gate (𝑂𝑂𝑡𝑡 ), and forgetgate ( 𝐹𝐹𝑡𝑡 ). The variations of LSTM networks may notnecessarily include all the mentioned gates. Instead, they mayhave other gates suited for the specific related applications.