Electricity consumption forecasting using neural networks for low-carbon power systems planning

E3S Web of Conferences 351, 01069 (2022) ICIES’22 https://doi.org/10.1051/e3sconf/202235101069 Electricity consumption forecasting using neural networks for low-carbon power systems planning Daniel Perez-Moscote,1,*, and Mikhail Tyagunov.2 1 National 2 National Research University "MPEI", Moscow, Russian Federation Research University "MPEI", Moscow, Russian Federation Abstract. Power systems require the continuous balance of energy supply and demand for their appropriate functioning, which makes electricity forecast a necessary process for the successful planning of operation and expansion of modern power systems, especially with the increase of renewable energy resources to be accommodated in order to realize low-carbon power systems. The task of predicting electricity consumption is complex because electricity demand patterns are intricate and involve various factors such as weather conditions. Recurring Neural Networks (RNN), such as Long Short-Term Memory (LSTM) networks, can learn long sequence patterns and make multi-step forecasts at once considering several variables, which can be especially useful for time series forecasts such as electricity consumption. This paper presents the application and assessment of a multivariate multi-step times series forecasting model based on LSTM neural networks for short-term prediction of electricity consumption using a dataset that encompasses data on energy load and meteorological elements from Belgorod Oblast in Russia as a case study. 1 Introduction Forecasting electricity consumption is necessary to adequately balance supply and load demand in modern power systems and, therefore, to plan operations and infrastructure expansion [1], especially in the context of the current transformation of the energy sector aiming at realizing low-carbon or carbon-neutral power systems with high penetration of variable renewable energy and large electrification of energy loads, as a major strategy of climate change mitigation. The appropriate forecast of electricity consumption in power systems is essential for the effective application of demand response programs and demand-side flexibility strategies to balance supply and demand allowing higher participation of variable renewable generation in the energy mix. The prediction of electricity consumption can be defined as time series forecasting problem [2] where the complexity of temporal dependence between observations is considered and the estimation of future loads is made based on previous data. Conventional time series forecasting methods focus on single-variable data with linear relationships and static dependence [3]; neural networks offer the capability to learn and approximate arbitrary non-linear functions supporting multivariate and multi-step forecasting [4], and especially Recurrent Neural Networks (RNN) could handle ordered observations and learn time-dependent context. In this order, Long Short-Term Memory (LSTM) networks, a type of RNN, are claimed to be able to automatically learn the characteristics and patterns of the time series dataset, supporting multiple variables to generate multi-step predictions [3]. To confirm this assumption, we applied and assessed a time series forecasting model based on LSTM for the prediction of electrical energy consumption using past sequences of consumption data as well as past weather conditions records. A set of real data from Belgorod Oblast, in Russia, is used as a case study and this paper presents the results obtained. 2 Methodology To evaluate the suitability and performance of RNN for power consumption prediction, an LSTM forecasting model [5] has been adapted and applied to a dataset of actual power consumption and weather conditions over a year from the Belgorod Oblast to make predictions on consumption for the following 24 hours using the previous 72 hours. The model is developed using Python and Keras, a deep learning application programming interface (API) that runs on top of the TensorFlow machine learning platform [6]. This is a multivariate multi-step time series forecasting problem and it is defined by the following function: [𝑌𝑌𝑡𝑡, 𝑌𝑌𝑡𝑡+1, …, 𝑌𝑌𝑡𝑡+𝑛𝑛−1] = 𝑓𝑓 (𝑌𝑌𝑡𝑡−1, 𝑌𝑌𝑡𝑡−2, …, 𝑌𝑌𝑡𝑡−𝑝𝑝, 𝑋𝑋𝑡𝑡−1, 𝑋𝑋𝑡𝑡−2, …, 𝑋𝑋𝑡𝑡−𝑝𝑝) (1) where Y is a variable to be predicted n steps (hours) ahead using data from the previous p hours. In this case, Y is the power load; X represents the weather variables * Corresponding author: © The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 (http://creativecommons.org/licenses/by/4.0/). E3S Web of Conferences 351, 01069 (2022) ICIES’22 https://doi.org/10.1051/e3sconf/202235101069 that influence the value of Y, in this case, air temperature, atmospheric pressure, humidity, and day length; n is 24 hours, and p is 72 hours. There are several stages in our time series forecasting analysis: 1. Data visualization and preparation for the model 2. LSTM model definition and fitting 3. Model training and testing 4. Future forecasting beyond the dataset 2.1. Data visualization and preparation Belgorod Oblast is a constituent entity of the Russian Federation with an area of 27134 km2. The power load dataset was collected every hour from 2020.20.04 to 2021.03.31 (working days). Data on weather characteristics were collected at the meteorological station located at Belgorod International Airport [7], 4 km north of Belgorod city, the administrative center of Belgorod Oblast. The dataset includes the hourly past values of air temperature (°C), atmospheric pressure (mmHg), and humidity (%). Lastly, the dataset also integrates day length, in minutes, for each day for the city of Belgorod [8]. There are a total of 5566 data points and Figure 1 shows the first and last rows of the dataset. Fig. 2. Hourly power load for the past year – Belgorod Region. Fig. 3. Daylength in minutes for the past year - Belgorod Region. The dataset must first be divided into a training set and a testing set. 90% of the datapoint will be used for training the model, and then the model should forecast the remaining 10% to be compared with the real values in the testing set and evaluate the performance. Fig. 1. Electrical load, air temperature, atmospheric pressure, humidity, and daylength dataset - Belgorod Region. Figure 2 presents all the data points representing the hourly power load throughout the year, and Figure 3 shows the data points representing day length. Introducing the daylength dataset into the forecasting model, allows the system to consider the seasonal fluctuations in electricity demand, as demand is generally higher in winter than in summer. 2 test_share = 0.1 # testing share: 10% of dataset The input to an LSTM model is a three-dimensional array: (Samples, Timesteps, Features).  Samples are the total number of sequences built for training.  Timesteps is the time length of each sample.  Features are the n (...truncated)