The growing demand and improvements in manufacturing capabilities, supported by government subsidies, has allowed the increase in the installed capacity of utility scale solar parks. Due to the remoteness in their location, the costs associated with dispatching personnel for maintenance is extremely high. A major contribution towards unscheduled downtime of these plants is due to the inverter faults. Currently, reactive and preventive maintenance are the most prevailing methods to identify and fix inverter faults. The presumption that the components will not under-performor fail until the scheduled visit, leads to a significant loss of production and revenue. To deal with the disadvantages of current maintenance methods, the solar industry is very keen on understanding the possibility of early detection of inverter faults by implementation of predictive maintenance. This research assessed the applicability of Machine Learning (ML) towards early signal detection of inverter faults in order to generate predictive maintenance alerts. The data for building the ML algorithms was acquired from a Shell owned 26.6 MWp utility scale solar park located in Moerdijk, The Netherlands. The early signal detection algorithmdeveloped, was based on the comparison between the actual and the predicted active power. The model built to predict the active power was based on two supervised learning methods;Elastic Net and Gradient BoostingMachine (GBM) with quantile regression. These models were capable of predicting the active power with a Mean Absolute Error (MAE) of 0.98kW & Root Mean Square Error (RMSE) of 1.8kW using Global Plane of Array irradiance (GPOA) and module temperaturemeasurements available from theMoerdijk data. The early signal detection relied on differentiating between prediction error and actual error. A window was created to encompass the maximum extent of prediction errors to avoid any false positive signals. This window for elastic net was found to be ¡¾ on the lower side and 2¾ on the upper side. Although when elastic net method was tested on 337 inverters- by looking at their residual variation 1-week prior to registered fault- it was found that the predictions suffered a periodic structural error. This was due to the erroneous predictions at times with extreme irradiance values. To mitigate, this the GBM with quantiles of 0.01 and 0.99 of GPOA was built to create a range of predictions giving rise to a wider range for normal operation. The results from both the algorithms indicated no early signals for inverter fault detection. This was partly due to data quality issues with fault tags in the Supervisory Control and Data Acquisition (SCADA) monitoring system; only 7 actual fault cases were identified. Additionally, the economic feasibility of implementing predictive maintenance was found to potentially reduce the current Operational Expenses (OPeX) by up to 10%. Despite the issues with data quality, an approach of using ML towards early fault detection for inverters in utility scale solar parks has been realised through this research.