LSTM auto-encoder based representative scenario generation method for hybrid hydro-PV power system

Abstract

The increasing penetration of renewable energy sources causes complex uncertainties of the power system. To capture such uncertainties in power system planning, an important step is to generate representative scenarios. In this work, a long short term memory (LSTM) auto‐encoder based approach is proposed to generate representative scenarios in an integrated hydro‐photovoltaic (PV) power generation system, which consists of feature extraction by LSTM Encoder, scenario clustering in feature domain by combining gap statistics method and K‐means++, and representative scenario reconstruction by using LSTM Decoder. Compared with traditional scenario selection and generation methods, the proposed method can better capture the patterns of multivariate time‐series data in both temporal and spatial dimensions. A case study in southwest China is used to demonstrate the effectiveness of the proposed method, which outperforms other existing methods by achieving the lowest SSE and DBI indices of 0.89 and 0.12, respectively, and obtaining the best SIL and CHI scores of 0.93 and 2.30, respectively, In addition, the case study shows the proposed model setup works more stable for scenario generation.

Full text

IET Generation, Transmission & Distribution
Special Issue: Advanced Data-Analytics for Power System Operation,
Control and Enhanced Situational Awareness
LSTM auto-encoder based representative
scenario generation method for hybrid hydro-
PV power system
ISSN 1751-8687
Received on 21st April 2020
Revised 23rd June 2020
Accepted on 14th July 2020
E-First on 4th August 2020
doi: 10.1049/iet-gtd.2020.0757
www.ietdl.org
Jingxian Yang1, Shuai Zhang1, Yue Xiang1,2 , Jichun Liu1, Junyong Liu1, Xiaoyan Han3, Fei Teng2
1College of Electrical Engineering, Sichuan University, Chengdu 610065, People's Republic of China
2Department of Electrical and Electronic Engineering, Imperial College London, London SW7 2AZ, UK
3State Grid Sichuan Electric Power Company, Chengdu 610041, People's Republic of China
E-mail: xiang@scu.edu.cn
Abstract: The increasing penetration of renewable energy sources causes complex uncertainties of the power system. To
capture such uncertainties in power system planning, an important step is to generate representative scenarios. In this work, a
long short term memory (LSTM) auto-encoder based approach is proposed to generate representative scenarios in an
integrated hydro-photovoltaic (PV) power generation system, which consists of feature extraction by LSTM Encoder, scenario
clustering in feature domain by combining gap statistics method and K-means++, and representative scenario reconstruction by
using LSTM Decoder. Compared with traditional scenario selection and generation methods, the proposed method can better
capture the patterns of multivariate time-series data in both temporal and spatial dimensions. A case study in southwest China
is used to demonstrate the effectiveness of the proposed method, which outperforms other existing methods by achieving the
lowest SSE and DBI indices of 0.89 and 0.12, respectively, and obtaining the best SIL and CHI scores of 0.93 and 2.30,
respectively, In addition, the case study shows the proposed model setup works more stable for scenario generation.
1 Introduction
Power generation using renewable energy sources gaining
momentum among power sectors in recent years owing to the fast
depletion of fossil fuels and consideration to reduce greenhouse
gases emissions [1]. Of all renewable sources, solar energy has
achieved significant market share and been recognised as a
competitive resource because of its simple construction process,
declining prices, and low operating costs [2]. As the ‘Renewable
Energy Statistic’ stated, by the end of 2019, the global solar
photovoltaic (PV) installation reached 580 GW, a 97.6 GW annual
increase and accounting for nearly 55% of newly increased
renewable energy capacity in 2019. However, the randomness and
intermittency nature of PV generation affects the security and
reliability of the power system. A promising approach is to operate
PV generation together with other renewable energy resources as
hybrid energy systems (HESs) to provide greater balance in energy
supply [3, 4]. The hybrid hydro-PV system is among the most
widely used HES as hydropower can be regulated rapidly to
complement solar energy [5]. Along with the successful
commissioning of the world's largest complementary hydro-PV
plant-Longyangxia hydro-PV plant (850 MW PV arrays + 1280
MW hydropower units) – in northwestern China in 2013,
increasing hydro-PV systems are being developed in regions that
are rich in both hydro and solar resources, such as west China [6–
8]. However, hydro-PV power generation may not completely
follow the local load demand, leading to a growing probability of
load shedding and reduced productivity of hybrid power plants. So
with the increase in penetration of PV resources, it is important for
system planner and operator to analyse and manage the uncertainty
of PV power generation to maintain the balance of integrated
hydro-PV power system and accommodate the expected load
increasing over the planning horizon [9–11].
The economic planning in power system involves the analysis
and evaluation of different scenarios and conditions. The multi-
scenario operating and planning models have been widely used to
capture system operating variations in power system planning [12–
14]. Solar radiation, hydropower and load may display daily or
seasonal patterns over longer time periods, although they follow a
stochastic process in the short term [15]. Therefore, generating
scenarios that characterise the uncertainty and reflect the periodic
variation of regional load, PV power generation and hydropower
generation is an essential step in the planning of hybrid hydro-PV
power system [16].
There are two widely used methods for scenario generation: (i)
probabilistic distribution based approach [17, 18]; and (ii) data
mining-based approach. The former approach mainly follows two
steps, which includes obtaining the probabilistic distribution of
power generation or errors and sampling the scenarios from the
statistical distribution. For example, Diaz et al. [19] established the
cumulative distribution function of wind power and generated
scenarios with Latin Hypercube Sampling method. The scenario
for wind units and cascaded hydro generation is generated with the
Monte Carlo method via the statistical distribution of forecast
errors in [20]. The wind speed scenarios are formulated with hourly
time series data based on the transition matrix approach of the
Markov Chain process [21]. However, large-scale power systems
usually involve different generating components and loads, it is
extremely challenging to develop a general probabilistic model that
captures both spatial–temporal and across-component correlations.
For the latter approach, clustering analysis, as an unsupervised data
mining technique, has been applied to efficiently reflect the
diversity of scenarios [22]. Baringo and Conejo [23] adopted K-
means clustering to characterise the scenarios of electric load and
wind production for investment decisions. Different clustering
techniques, such as Fuzzy C-Means, Gaussian mixture models and
Dynamic Time Warping (DTW), have been applied in daily
clearness index profiles to obtain typical scenarios of PV power
generation [24]. Hierarchical clustering and density-based
clustering are also adopted to cluster load profiles to recognise
different underline patterns [25, 26].
Traditional clustering algorithms face significant challenges
when applied to multivariate time sequence hydro/PV/load data: (i)
traditional clustering methods, which often cluster single variable,
are lack of effective uniform similarity measurement structure for
multivariate hydro/PV/load data; (ii) massive and high dimensional
data may cause complexity and heavy computational burden; (iii)
the spatial and temporal correlations between uncertain
hydropower, PV and load cannot be fully captured by traditional
clustering technologies [27, 28].
IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 24, pp. 5935-5943
© The Institution of Engineering and Technology 2020
5935