Neural Network AI Techniques
In Ref. [19], a method of forecasting is proposed for the analysis of data obtained from a PV array. This method incorporates long-short term memory (LSTM) and CNN techniques together. This method can even provide the forecast for PV farms for the following day on whether some sensors are not in working condition or are not installed. In this methodology, selection of sufficient layers was done for the regressing the upcoming day’s solar power. This was done on the basis of the previous day’s data, i.e. values of irradiance and the panel’s temperature taken every 10 minutes. The proposed method utilized the approximate data gathered from the meteorological centre and improved the performance of the PV system. In Ref. [20], a deep NN model is presented for accurate forecasting of a PV system’s output power. It can produce a one-day forecasting of output power on the basis of input data such as irradiance, temperature, and past values of output power of PV systems. The simulation results showed that the proposed scheme was able to outperform other models in terms of accurate forecasting when subjected to highly irregular and unstable input data.
In Ref. [21], a conventional NN method is employed for pre-processing of EL images and dividing them into cells that serve as input data for ML algorithms. The pre-process of raw EL image data requires correction in lens distortion, filter and threshold operation, regression fitting, and transformation. Conventional NN performed better in terms of classification of PV cells as compared to random forest technique and SVM. In Ref. [22], an ANN method is proposed for predicting one day in advance irradiance curve under different meteorological conditions. This method improves the unprocessed forecast of numerical weather prediction by merging it with statistics-based learning schemes. The method was able to reduce the uncertainty in electrical power generation of PV systems. In Ref. [23], a forecasting method is presented for sun irradiance that overcame the limits of the LSTM method. This method encodes the time-series information into image form by combining LSTM with Gramian angular field. The advantage of this method is that it requires a small data set but provides excellent forecasting of irradiance.
In Ref. [24], a method for computation of global horizontal irradiance (GHI) has been proposed. Conventional methods for computing GHI involve costly equipment, workstations, and satellite-based models. Compared with traditional methods, this method is more accurate and convenient for prediction of GHI. In this work, CNN was applied on regression of images and detection and elimination of anomaly data was done using Gaussian and Bayesian models. The processed input having 3-month images and GHI data were fed into the proposed approach. Statistical analysis showed that the CNN-based method was more accurate, swift, inexpensive, and convenient for monitoring of big farms as compared to traditional irradiance measurement approaches.
The operational life of PV system can be increased if the diagnosis of fault is made properly. There are several conditions encountered such as shading, impedance of large value, mismatch of location, MPP tracking, and adverse weather that make the detection of faults difficult. In Ref. [25], a deep two-dimensional CNN is proposed for the extraction of features from 2-D graphs obtained from data of PV systems. The presented methodology achieved accuracy of approx. 74% for the detection of fault and approx. 70% for a noisy scenario. It is important to forecast a day advance for PV plants as it is desired for economic dispatch of power, energy management, and commitment of units. In Ref. [26], a similarity approach-based forecasting method is proposed for the prediction of PV output power in high resolution using low-resolution variables of weather. Their results demonstrated that the proposed methodology provided greater accuracy as compared to other forecasting models.
In Ref. [27], a method is proposed for short period scheduling of line maintenance in PV systems connected to a distribution network. It is well known that PV output power shows intermittent behaviour due to the involvement of random or fuzzy input variables that makes its prediction difficult. In the proposed strategy, initially, atmospheric data like irradiance and the amount of cloud cover are gathered. In the next step, uncertainties and scheduling of the distribution network are modelled with the help of a fuzzy program. It follows a pessimistic approach and optimizes the unenthusiastic values from the perspective of reliable and economic operation when subjected to probability constraints. In the last stage, a hybrid AI method was applied for finding the solution of the model. IEEE 33-Bus system was chosen for simulations and experiments. The outcomes of the proposed work validate the efficacy of the presented line maintenance for short-term scheduling of distribution network when subjected to uncertainties of fuzziness and randomness.
Adaptive Neuro-fuzzy interference system (ANFIS) is another powerful AI tool used for fast and dynamic MPPT control. In Ref. [28], ANFIS is applied for improving the accuracy and response of MPPT in standalone PV systems. The parameters controlled using the proposed scheme were injected power, severe value of voltage, frequency, and current. In Ref. [29], the role of internet of things (IoT)-based wireless sensor networks (WSN) in MPPT of PV arrays was discussed. The use of WSN in PV systems can reduce the requirement for complex hardware significantly, which results in low cost. In Ref. [30], an adaptive structure model is developed for forecasting a day in advance accessible energy in PV systems. The developed framework used a data analysis method that combined statistical and AI techniques.
In Ref. [31], the combined use of IoT with ANN and ANFIS is proposed for the prediction of power generated from PV systems. In terms of performance, ANN- based IoT is superior to ANFIS. In Ref. [32], AI was used for development of smart energy systems in cities. These studies all concluded that AI can help in improving the output of PV systems and in increasing the productivity of labour, which can pave the way for smart cities. In Ref. [33], an ANN-based model was implemented for controlling and dispatching steam power generation using a solar field. In this study, an ANN model having two BPs with four hidden layer networks was developed. The model utilized the prior inputs obtained from the steady-state heat transfer model. The first hidden 4 layers model the data on the solar field’s temperature and pressure, while the second hidden layers perform analysis on the solar field’s outlet temperature. The findings of this work suggest that the error in the prediction of pressure and temperature were reduced as compared to that obtained from the BP model and RNN while accuracy in simulation increased.
Management of energy is an essential goal to achieve in the development of a smart grid. This is because of integration of intermittent RE resources into the grid. Hence, the use of AI in smart grid is proposed to accurately forecast the generation from RE systems, including PV plants. In Ref. [34], a hybrid adaptive learning model is proposed which could make accurate predictions about the intensity of solar irradiance on the basis of weather data. In this research, a BP neural network model incorporating genetic algorithm was presented and applied on learning nonlinear associations in data. The presented methodology was able to detect the time-based, linearity/nonlinearity relationships in data, which resulted in enhancing predication ability. In simulation results, it was shown that presented methodology was superior to its counterparts in forecasting of solar intensity on a long-term or short-term basis. AI also plays an important role in the control and design of PV systems. In Ref. [35], recent AI methods used in the prediction of output power, diagnosis of faults, and controlling and calculation of optimal size in PV systems are reviewed. In the review, it was found that AI techniques such as artificial immune system (AIS), bee colony, and artificial fish swarm algorithm (AFSA) were more accurate in terms of identifying simulated annealing (SA) and GA. In a survey, NN was found to be superior in terms of accurate forecasting for MPPT and sizing PV systems and also in the diagnosis of fault.
In recent years, there has been bulk deployment of smart meters at the consumer end. The information gathered from these smart meters can be utilized for the forecasting of load demand. It can also improve the pattern of energy consumption adopted by consumers by using energy management methods. However, along with the many merits of smart meters, there are several challenges reported about this scheme. Data obtained through smart meter are volatile, variable, complex, and of large size. Management of this type of data is quite difficult. In Ref. [36], a new methodology is proposed based on clustering of data obtained through smart meter. This approach was presented for the fulfilment of the following objectives:
- 1. To fine detailed profile of load
- 2. To reduce complexity in load profile
- 3. To forecast load demand accurately
- 4. For optimum management of energy
In the presented concept of clustering, the data size of the smart meter was reduced. The simulation results validate that forecast of load was improved significantly using the clustering algorithm. In addition to this advantage, it also lightened the burden of mathematical calculation. It was also shown that the management of energy was improved as the proposed scheme resulted in larger savings of cost.
