ARTIFICIAL INTELLIGENCE BASED MODELS FOR SHORT-TERM SOLAR IRRADIANCE FORECASTING

Abstract

The ever-growing population, their dependence on electricity, and awareness of the tremen dous environmental impact of burning fossil fuels has expanded the renewable energy portfolios across the globe. Solar energy is arising as the most promising alternative to the fossil fuels. Among all the techniques that transforms solar energy to electricity, solar photovoltaic (PV) is becoming the most popular due to its simplicity and cheap maintenance. The recent advance ments in PVtechnologies has rocketed its installation and efficiency in commercial applications along with a significant reduction in their costs. Although the solar energy has the advantage of being limitless and clean over conventional resources, it brings along several challenges. The PV power output is highly volatile as it depends on several meteorological factors, including solar irradiance, temperature, cloud cover, rainfall, etc. Solar energy is also an intermittent energy source as it only exists during day time. The uncertain and intermittent nature of the solar energy is the main hindrance in its reliable market penetration. The variability of solar power output affects the grid balance system and increases their operational costs. Therefore, with the increased installation of PV plants across the globe, accurate forecasting models are highly desirable for the successful integration of solar energy to the grid and proper functioning of energy industry. The solar resource availability of a location depends on the local meteorological parameters. As using a large number of input parameters for model development increases the computa tional cost unnecessarily, it is beneficial to identify the most influential and useful parameters for model development. Therefore, the selection of suitable meteorological variable is necessary to develop the accurate and efficient solar irradiance prediction model. In our first objective, a case study of mountainous region of Uttarakhand, India is conducted to determine the solar potential of Uttarakhand state using the most influential input parameters. To perform this, ar tificial neural network (ANN) based global horizontal irradiance (GHI) prediction models are developed with different combinations of meteorological parameters, which includes minimum temperature (Tmin), maximum temperature (Tmax), temperature difference (∆T), GHI, extrater restrial radiation (H0), and bright sunshine hours (S). The results revealed that two to three input parameters can efficiently estimate the GHI in Uttarakhand. Moreover, the performance of de veloped models is also validated with the measured and satellited based data. Further, the ANN model with best input parameters is used to assess the solar potential of thirteen districts of Uttarakhand. Machine learning techniques suffer from several limitations, including high computational cost, instability issues, and less accuracy while handling high-dimensional and complex data. In this context, a more advanced approach called ‘ensemble learning’ for solar irradiance forecast is utilized in our second objective. In this, a new ensemble model, which consists of two ad vance base models, namely extreme gradient boosting forest and deep neural networks (XGBF DNN), is proposed for hourly global horizontal irradiance forecast. To avoid the problem of over-fitting, the base models are integrated using ridge regression. Moreover, the proposed framework is designed carefully to ensure the diversity in base models, as diversity among base models is widely recognized as a key to the success of an ensemble model. Further, to en hance the model’s accuracy, a subset of input features is selected, which includes temperature, clear-sky index, relative humidity, and hour of the day as the most relevant features. To provide a comprehensive and reliable assessment, proposed model is validated with data from three different climatic Indian locations, including New Delhi, Jaipur and Gangtok. Subsequently, a seasonal analysis is also carried out to provide a deeper insight into the model’s performance. The performance of the proposed model is evaluated by comparing the prediction results with different models, including benchmark smart persistence and traditional machine learning techniques, such as random forest, support vector regression, extreme gradient boosting forest and deep neural networks. The proposed ensemble model exhibited the best combination of stability and prediction accuracy irrespective of seasonal variations in weather conditions. The achieved predictive performance and stability of the proposed model made it an ideal and reliable recommendation for hourly GHI prediction. With the countless advancements in computer software, hardware and availability of huge datasets or big data techniques, deep learning techniques are steadily emerging. Therefore, the development of a deep learning based solar irradiance forecasting models is highly desirable to further enhance the prediction accuracy. Moreover, the distribution and availability of solar irradiance at a region majorly depends on the climatic conditions and latitude of that location.The meteorological data of a location along with that of near by locations is very helpful in predicting the solar irradiance of that location. Moreover, the temporal information in the time series data of GHI is also very informative for GHI prediction. In this way, both spatial and temporal features can be utilized in GHI forecasting. To address this, a deep hybrid LSTM-CNN model, which integrates long short-term memory (LSTM) with convolutional neural network (CNN) to model spatio-temporal features for short-term GHI forecasting, is proposed in our third objective. The proposed model is trained with the meteorological data of 23 locations of California State, USA, which includes temperature, clearness index, relative humidity, cloud cover, etc. as input parameters. The proposed hybrid LSTM-CNN model firstly uses LSTM to extract the temporal features from time series GHI data followed by CNN which extracts the spatial features from the correlation values between several meteorological variables of target and its neighbour location. The prediction accuracy of the developed model is analysed rigorously by examining the performance for an year, for four seasons and under three sky conditions. The findings of the study suggest that the proposed hybrid LSTM-CNN models is a reliable alternative for short-term GHI prediction due to its high predictive accuracy and stability under diverse climatic, seasonal and sky conditions.