Relatively little attention has been given to the short-term load forecasting problem of primary substations mainly because load forecasts were not essential to secure the operation of passive distribution networks. With the increasing uptake of intermittent generations, distribution networks are becoming active since power flows can change direction in a somewhat volatile fashion. The volatility of power flows introduces operational constraints on voltage control, system fault levels, thermal constraints, systems losses and high reverse power flows. Today, greater observability of the networks is essential to maintain a safe overall system and to maximise the utilisation of existing assets. Hence, to identify and anticipate for any forthcoming critical operational conditions, networks operators are compelled to broaden their visibility of the networks to time horizons that include not only real-time information but also hour-ahead and day-ahead forecasts. With this change in paradigm, progressively, large scales of short-term load forecasters is integrated as an essential component of distribution networks' control and planning tools.

The data acquisition of large scale real-world data is prone to errors; anomalies in data sets can lead to erroneous forecasting outcomes. Hence, data cleansing is an essential first step in data-driven learning techniques. Data cleansing is a labour-intensive and time-consuming task for the following reasons: 1) to select a suitable cleansing method is not trivial 2) to generalise or automate a cleansing procedure is challenging, 3) there is a risk to introduce new errors in the data. This thesis attempts to maximise the performance of large scale forecasting models by addressing the quality of the modelling data. Thus, the objectives of this research are to identify the bad data quality causes, design an automatic data cleansing procedure suitable for large scale distribution network datasets and, to propose a rigorous framework for modelling MV distribution network feeders time series with deep learning architecture. The thesis discusses in detail the challenges in handling and modelling real-world distribution feeders time series. It also discusses a robust technique to detect outliers in the presence of level-shifts, and suitable missing values imputation techniques. All the concepts have been demonstrated on large real-world distribution network data.