Stochastic Synthetic Data Generation for Electric Net Load and Its Application M. Vivienne Liu Cornell University ml2589@cornell.edu Patrick M. Reed Cornell University patrick.reed@cornell.edu C. Lindsay Anderson Cornell University cla28@cornell.edu Abstract The increasing integration of renewable energy in electric power systems focuses attention on realistic representation of “net load” because it aggregates the information from both demand and the renewable supply side; net load is the remaining demand that must be met by non-renewable resources. However, the net load data is not readily accessible because of cost, privacy and security concerns. Furthermore, even if historical data is available, multiple stochastic scenarios are often needed for a wide range of power system applications. To address these issues, this paper proposes a stochastic synthetic net load profile generation approach. A seasonal detrending technique is combined with the modified Fractional Gaussian Noise method to deal with the complex multi-periodic seasonal trends in the net load profile. A thorough statistical validation and temporal correlation check are performed to show the quality of the synthetic data. The benefits of the synthetic data are demonstrated by a microgrid energy management problem. 1. Introduction The system load profile is essential to a wide range of power system management applications including energy resource planning, electricity market clearing, risk assessment, reliability analysis, and policy design [1, 2]. However, as the effects of climate change intensify, integration of intermittent renewable energy resources is accelerating[3, 4]. Under these conditions, the load profile is insufficient to readily inform most power system operation and planning applications. As a result of the increasingly distributed nature of renewable energy resources, they are often consumed or stored locally (behind the meter) and invisible to the system operator [5]. Thus, instead of using load profile, net load profile, which is defined as the difference of load and any behind-the-meter energy, is increasing in importance[6]. Net load profile is not easily accessible by researchers and the reason is two-fold; first, it is expensive to install and maintain ubiquitous equipment necessary to collect electricity demand and renewable generation data with high spatial and temporal resolution; second, renewable generation is relatively nascent at large scale, and as result, historical data is scarce (for example, only 3 years’ data is available from CAISO) [7]. Thus, a novel modeling strategy to capture the stochastic time-varying behavior of both the electricity demands and the renewable supply would fill an important need in power systems planning and operations research. In addition to lack of availability of historical net load data, the growing interest in stochastic optimization and sequential time-series solutions [8, 9], and approximate dynamic programming, policy-based decision making approaches, [10, 11, 12, 13] requires a large set of plausible scenarios of net load input data to inform and enhance the decision making process. The set of stochastic net load profiles must preserve the statistical properties and temporal correlation of historical records [14]. Significant attention has been paid to time-varying synthetic renewable energy profile generation [15, 16, 17, 18, 19, 20, 21, 22, 23]. The Markovian state transition property has been an underlying assumption of the renewable energy behavior in much of this literature; first and second order Markov Chain-based methods have been proposed by [16, 17, 19, 20, 21, 23]. Most of these methods were able to preserve the probability distribution of the historical records, but temporal correlation has only been considered in [16, 20, 20] though these efforts leave a significant gap between the synthetic and historical data. Authors in [22] used a Fourier series and auto-regressive moving average model to capture the characteristics of historical data. The synthetic profiles showed promising performance in quantile and cumulative density function validations, though the temporal correlation was not explored as a key feature for time-series data generation. Proceedings of the 54th Hawaii International Conference on System Sciences | 2021 Page 3147 URI: https://hdl.handle.net/10125/70998 978-0-9981331-4-0 (CC BY-NC-ND 4.0)