Collaborative Real-Time Operation for Long-Term and Short-Term Energy Storage: A Scheduling-Corrective Approach

Under the background of carbon peaking and carbon neutrality, the renewable-dominated power grid attracts wide attention. To address the fluctuations of renewable power in different timescales, the collaborative operation of long-term and short-term energy storage devices is required. This paper studies the year-round operation of a renewable-dominated power grid. A data-driven scheduling-correction framework is proposed, which consists of offline scheduling and real-time operation. The scheduling module generates optimal storage level sequences of long-term energy storage in previous years ahead of the operating year using historical data of renewable power. The real-time module performs two stages in each period: The reference storage level is first determined for long-term storage based on historical optimal curves and newly observed renewable generation data; then a bi-objective rolling-horizon optimization problem is solved, in which the current operating cost is minimized and the long-term storage is steered to the reference value. An appealing feature is that the proposed framework utilizes only short-term prediction of renewable power. Numerical tests show that the proposed method outperforms existing rolling horizon approach in terms of operating cost, load shedding and renewable spillage rate. The collaboration of the long-term and short-term storage is also well-achieved.

I. Introduction

In order to achieve sustainable development, renewable energy has been gradually integrated into the power system [1], [2]. Power balance is a key issue in power systems with a high penetration of renewable energy. In daily operation, the unit commitment problem or economic dispatch problem are formulated to handle the intra-day power balance. In long-term operation, the system has to cater for the seasonal variations of both renewable power and load demand. However, the annual variation of available renewable power is generally significant, and it is hard to maintain the long-term power balance relying on conventional units. Energy storage devices are effective solutions to mitigating the power fluctuations, which have been extensively used in the safe operation and optimal control of power system, such as frequency regulation [3] and economic dispatch [4]. The incorporation of energy storage devices has become an indispensable strategy for the evolution of renewable-dominated power systems [5]. However, a singular type of storage often falls short of addressing the power fluctuations in different time scales. Battery stor-age is a representative of short-term storage. It has a high charging/discharging efficiency, but the self-discharging effect cannot be ignored in long-term operation. In contrary, the long-term energy storage, such as the pumped hydro storage, hydro-gen storage and compressed air energy storage, generally has a neglectable self-discharge rate, but the charging/discharging efficiency is low compared to battery. In order to address the frequent charging and discharging demands in short-term operation and meet the energy reserve requirements in long-term operation, the collaboration of long-term and short-term energy storage is of extreme importance. There have been various studies on scheduling and operation in renewable-dominated systems with energy storage devices, which can be classified into two categories. The first category adopts two-stage optimization methodology. Stochastic programming [6] and robust optimization [7] are widely used in the existing studies. Stochastic programming generally assumes known probability distribution of uncertainty and optimizes the ex-pectation. In robust optimization, the random variables are restricted in a bounded uncertainty set whose size is to do with some confidence level. Stochastic programming needs large number of sample scenarios, and robust programming requires uncertainty sets of renewable power. Therefore, the above methods need adequate historical data, which are generally conducted in daily operation. In long-term operation, the two- stage framework is hard to be adopted since the computation complexity will be huge.