On autoregressive deep learning models for day-ahead wind power forecasting with irregular shutdowns due to redispatching
Stefan Meisenbacher, Silas Aaron Selzer, Mehdi Dado, Maximilian Beichter, Tim Martin, Markus Zdrallek, Peter Bretschneider, Veit Hagenmeyer, Ralf Mikut · 2024
Read On autoregressive deep learning models for day-ahead wind power forecasting with irregular shutdowns due to redispatching by Stefan Meisenbacher, Silas Aaron Selzer, Mehdi Dado, Maximilian Beichter, Tim Martin, Markus Zdrallek, Peter Bretschneider, Veit Hagenmeyer, Ralf Mikut (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-00423.txt.
Introduction
Stefan Meisenbacher and co-authors study day ahead wind power forecasts in a 2024 paper. Grid operators need those forecasts to plan backup actions as wind replaces older plants. Turbines are sometimes shut down in irregular grid interventions, and those gaps can distort forecasts that learn from past output.
The team compare three deep learning methods that use past output plus weather forecasts. They test them on data with regular and with irregular shutdowns. The paper maps where shutdowns hurt accuracy and which designs cope best. Reliable forecasts would let wind farms take a larger role in keeping the grid stable.
Catalog record
| CREATOR | Stefan Meisenbacher, Silas Aaron Selzer, Mehdi Dado, Maximilian Beichter, Tim Martin, Markus Zdrallek, Peter Bretschneider, Veit Hagenmeyer, Ralf Mikut |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-00423
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
On autoregressive deep learning models for day-ahead wind power forecasting with irregular shutdowns due to redispatching Stefan Meisenbacher, Silas Aaron Selzer, Mehdi Dado, Maximilian Beichter, Tim Martin, Markus Zdrallek, Peter Bretschneider, Veit Hagenmeyer, Ralf Mikut Abstract Renewable energies and their operation are becoming increasingly vital for the stability of electrical power grids since conventional power plants are progressively being displaced, and their contribution to redispatch interventions is thereby diminishing. In order to consider renewable energies like Wind Power (WP) for such interventions as a substitute, day-ahead forecasts are necessary to communicate their availability for redispatch planning. In this context, automated and scalable forecasting models are required for the deployment to thousands of locally-distributed onshore WP turbines. Furthermore, the irregular interventions into the WP generation capabilities due to redispatch shutdowns pose challenges in the design and operation of WP forecasting models. Since state-of-the-art forecasting methods consider past WP generation values alongside day-ahead weather forecasts, redispatch shutdowns may impact the forecast. Therefore, the present paper highlights these challenges and analyzes state-of-the-art forecasting methods on data sets with both regular and irregular shutdowns. Specifically, we compare the forecasting accuracy of three autoregressive Deep Learning (DL) methods to methods based on WP curve modeling. Interestingly, the latter achieve lower forecasting errors, have fewer requirements for data cleaning during modeling and operation while being computationally more efficient, suggesting their advantages in practical applications. Keywords deep learning, autoregression, data cleaning, redispatch, wind power curve, time series forecasting 1 Introduction Forecasting locally distributed Wind Power generation is required to prevent grid congestion by balancing the electrical transmission and distribution grids [ 1 ] . As Wind Power capacity expands in future energy systems, automating the design and operation of Wind Power forecasting models becomes inevitable to keep pace with Wind Power capacity expansion. However, two key challenges exist in developing such automated Wind Power forecasting models: First, scalable Wind Power forecasting models that achieve low forecasting errors are needed. On the one hand, offshore Wind Power farms, consisting of many centralized Wind Power turbines in a uniform terrain, lend themselves to holistic modeling approaches. Such models can incorporate the wind direction to account for wake losses (mutual influence of Wind Power turbines). In holistic modeling, the use of computationally intensive methods can be worthwhile, for example as in [ 2 ] using an autoregressive Deep Learning model. On the other hand, onshore Wind Power turbines are created decentrally in different terrains, e. g., in areas with little air turbulence like open fields or in the area of cities, forests, and hills with higher air turbulence. The diversity of onshore Wind Power systems, featuring turbines with varying characteristics and located in heterogeneous terrains, prompts questions about the cost - effectiveness of employing computationally intensive Deep Learning models for forecasting. Second, interventions in the Wind Power generation capabilities at regular and irregular intervals represent a challenge for designing and operating Wind Power forecasting models. Regular shutdowns comprise time - controlled shutdowns, such as those for bat protection and maintenance, while changing the operating schedule to prevent line overloads (redispatch) represents irregular shutdowns. 1 1 1 Other reasons for irregular shutdowns are sudden faults and the provision of balancing energy for frequency control. For planning redispatch interventions, the forecast is required to communicate the available Wind Power generation without incorporating any forecasts of future redispatch interventions. Hence, the preparation of the training data sub - set and model design must take this forecasting target into account. Despite the availability of various training data cleaning methods [ 3 , 4 , 5 ] , redispatch interventions can also impair the model operation in making forecasts if they rely on autoregression. More precisely, such methods make a Wind Power forecast based on a horizon of past target time series values alongside exogenous forecasts (future covariates) like wind speed and direction forecasts. Since redispatch - related shutdowns also appear in this horizon of past values, the model may also forecast future shutdowns, which has an undesirable impact on forecast - based planning [ 6 ] . Therefore, the present paper increases awareness of both challenges and compares autoregressive forecasting methods on data sets containing regular and irregular shutdowns. Three autoregressive methods are considered, namely Deep AutoRegression [ 7 ] , Neural Hierarchical interpolation for Time Series forecasting [ 8 ] , and the Temporal Fusion Transformer [ 9 ] , which are selected for comparison since they represent distinct state - of - the - art Deep Learning architectures. These autoregressive forecasting methods are compared with methods based on Wind Power curve modeling, which rely solely on weather forecasts and omit past values. For Wind Power curve modeling, the three Machine Learning methods eXtreme Gradient Boosting, Support Vector Regression and the MultiLayer Perceptron are considered. Additionally, the Original Equipment Manufacturer’s Wind Power curve and an Automated Machine Learning method based on Wind Power curves called AutoWP [ 10 ] are included. In contrast to other Automated Machine Learning methods [ 11 ] , AutoWP dispenses with the computationally expensive HyperParameter Optimization but uses computationally efficient ensemble learning, making it scalable for model deployment to thousands of individual Wind Power turbines. This article is an extension of the conference paper [ 10 ] and is organized as follows: related work is reviewed in Section 2 , applied methods are detailed in Section 3 , the evaluation is given in Section 4 and results are discussed in Section 5 , followed by the conclusion and outlook in Section 6 . 2 Related work This section first focuses on renewable energy forecasting before Wind Power forecasting methods are analyzed in the context of redispatch interventions. 2 2 2 Note that other applications, e. g., model - based observer and control design [ 12 ] , require different modeling methods. Finally, related work that considers inconsistent Wind Power data is reviewed and the research gap is summarized. As stated in the introduction, the growing share of renewable power generation results in an increasing demand for forecasts. In this context, PhotoVoltaic [ 13 , 14 , 15 ] and Wind Power forecasts [ 10 , 14 , 16 ] are strongly weather - dependent. For Wind Power forecasts, the methods can be separated into two types: 1. The first type comprises autoregressive methods, i. e., the forecast relies on the target turbine’s past Wind Power generation values. Weather forecasts can be incorporated as exogenous features. 2. The second type comprises methods that do not consider past values but instead model the so - called Wind Power curve, i. e., the empirical relationship between wind speed and Wind Power. Further additional explanatory variables can be taken into account. With regard to autoregressive methods (first type), methods based on Statistical Modeling, Machine Learning, and Deep Learning are proposed in the literature. Classical autoregressive Statistical Modeling methods include the Autoregressive Moving Average [ 17 ] and its integrated version, the ARIMA [ 18 ] , to cope with trends. Whilst they disregard the strong weather - dependency of Wind Power generation, it can be considered by extending Statistical Modeling with exogenous variables, e. g. ARIMAX [ 19 ] . Exogenous variables can also be considered in Machine Learning - based autoregressive