Data-driven Approaches for Power Grids

Modern power grids are characterized by the rapidly increasing penetration of fluctuating distributed energy resources (DERs). On the distribution level, reverse power flows and a greater ratio of locally fluctuating generation require the real time efficient and secure operation of power grids. However, distribution grids are poorly observable and would require tremendous investments in installing additional sensors for data collection. If it was possible to collect and use these data in real-time, the grid operators could determine the setpoint of optimal operation. Even if these data were available, however, with the millions of DER units connected to the grid, it is almost impossible to manage their operation centrally in real time. The communication and computation requirements for such a task go beyond the currently available capabilities of state-of-the-art communication infrastructure and computation devices. As a result, in the presence of no control, fluctuating DERs inevitably may cause critical events, increased power losses, and poor voltage profiles throughout the day. At the same time, algorithms that try to address issues such as minimizing the power losses by adjusting the power injection of inverter-interfaced DER, may compromise the inverters’ fault-ride-through capability. This can become crucial for the reliable operation of the distribution grids, as they could lose valuable resources to support grid voltage at the time they need them the most.

In this context, this thesis aims to develop data-driven solutions for the control, secure operation, and optimization of power grids. One of the research directions is the optimal design of Volt/VAR control curves for inverter-interfaced DERs. In order to demonstrate the practical applicability of the proposed control rules, constraints from the IEEE 1547.8 Standard are accounted for, which aims to provide secure integration and operation of DERs in the US grids. In addition to satisfying an extensive list of the IEEE 1547.8 constraints, the designed Volt/VAR control achieves to maintain a voltage level of around one per unit despite fluctuations of consumption and active power generation, and outperforms the existing methods. Additionally, in the proposed algorithm, rule parameters are customized per bus and optimized centrally by the operator on a two-hour basis. This allows to achieve a desirable voltage profile while guaranteeing the stability of the dynamical system.

For the systems with a fully unknown feeder model and millions of DERs, this thesis proposes two fully decentralized model-free loss minimization algorithms. The algorithms address the problem of active power loss minimization in distribution grids while keeping the voltage magnitudes on all buses within the defined limits. It is then analytically proven that the proposed algorithms are guaranteed to reduce grid losses without any prior information about the network, requiring no communication, and based only on local measurements. For the grey-box systems, i.e., with (partly) known topology and limited communication infrastructure, two hybrid algorithms are developed. The designed hybrid algorithms have much lower communication and computation requirements than traditional methods, while they also provide performance guarantees in the case of communication failure. Equally importantly, by carrying out extensive simulations on the real-time digital simulation (RTDS) platform, it is verified that the proposed algorithms comply with the grid codes for the low voltage ride through (LVRT) capability of the inverters. Additionally, it is shown that combining these algorithms with active power injection forecasting methods, such as the existing state-of-the-art Wavelet-CNN-LSTM, further enhances the performance of the developed local loss minimization algorithms.

Developing these optimization algorithms for distribution grids and combining them with neural networks for forecasting, two challenges were identified. First, neural networks still remain a black box when it comes to the transparency of their reasoning behind the generated output. Second, there are power grid stability constraints that are intractable for the conventional optimization algorithms; neural networks, though, are able to capture such constraints and, through an exact transformation, embed them in an optimization problem in a tractable way. To address the issue of neural network transparency, this thesis examines both regression and classification neural networks with the use of three gradient-based methods, which are able to interpret the neural network operation by connecting changes in the output with marginal changes in the input. These methods are implemented on neural networks that forecast aggregated photovoltaic (PV) generation, load, and their ratio to each other, given the real historical data. Moreover, we can further detect abnormal cases, when predictions might fail, by estimating the prediction uncertainty with Bayesian neural networks. Such neural networks, which are interpreted by the gradient-based methods and complemented with uncertainty estimation, provide robust and explainable forecasting for decision-makers. To address the second challenge and showcase how neural networks can capture previously intractable constraints and embed them in an optimization problem, this thesis develops a novel method that is applied to an AC optimal power flow (AC-OPF) problem with N-1 and dynamic security constraints (more specifically, small signal stability constraints). Leveraging an exact mixed-integer reformulation of neural networks, the whole security-constrained AC-OPF problem is cast as a mixed integer linear problem (MILP). The solution of the corresponding mixed-integer problem provides an accurate approximation to the originally intractable non-linear optimization problem and allows to efficiently obtain cost-optimal solutions, simultaneously satisfying both static and dynamic security constraints.