Advancing Ultra-Short-Term Wind Power Forecasting

with Multi-Channel ML Techniques

1st Jiapin Ren 1st Chenlin Zhu 1st Zeyu Huang
Department of Information and Metropolitan College School of Information Engineering
Electrical Engineering Boston University China Jiliang University
Xuhai College, China University of Boston , USA Hangzhou , China
Mining and Technology [email protected] [email protected]
Xuzhou , China
[email protected]

2nd Ruihan Chen 3rd Shurui Xu

School of Mathematical Sciences Cognitive Neuroscience Laboratory
Zhejiang Normal University Beijing Meishi Cha Technology Co., Ltd.
Jinhua, China Beijing , China
[email protected] *Corresponding author: [email protected]

4th Peng You

a
National and Local Joint Engineering Laboratory of Internet Applied Technology on Mines
China University of Mining and Technology
b
School of Information and Control Engineering ,China University of Mining and Technology
c
Xuhai College, China University of Mining and Technology
Xuzhou , China
*Corresponding author: [email protected]

Abstract―To address the inherent unpredictability of of renewable energy within the global energy matrix—
wind speeds and directions, which significantly affect wind particularly the swift advancement of wind power—precise
turbine output, this paper presents a multi-channel feature predictions of wind farm output in the ensuing minutes to hours
combination strategy for ultra-short term wind power have emerged as a critical research domain in grid scheduling
prediction. The research utilises a number of data sources, and energy management. Wind power uncertainty primarily
including meteorological observations, historical power arises from the highly unpredictable nature of wind speed and
data and geographical inputs, to create a robust feature set direction[2], which exert a direct and significant impact on
for prediction models. By using advanced machine learning turbine output. Minor variations in these meteorological
techniques, in particular deep learning models such as parameters can induce substantial power generation
fluctuations, and they are affected by diverse factors such as
Convolutional Neural Networks (CNNs) and Long Short-
topography, atmospheric conditions, and seasonal changes.
Term Memory (LSTM) networks, the study effectively
captures the complex dynamics of wind power generation. To enhance forecast precision and dependability,
The novel approach of reclassifying data sets and refining researchers have adopted a multi-channel feature combination
model features through task-specific channels significantly strategy. This method engenders a comprehensive input feature
improves prediction accuracy. Experimental results show set by amalgamating various data types, including
significant improvements in prediction accuracy, meteorological observations, historical power records, and
highlighting the potential of the proposed method to geographical as well as environmental considerations[3]. By
optimise grid operation and energy management in real- harnessing this extensive data integration, prediction algorithms
time scenarios. can more effectively apprehend the intricate dynamics
governing wind power production[4]. Nevertheless, the optimal
Keywords-Wind Power Forecasting; Feature Combination; selection and amalgamation of features, along with the handling
Energy; Management; CNNs,LSTMs; of noise and anomalies in voluminous datasets, remain daunting
challenges in refining the forecasting process.
State-of-the-art machine-learning methodologies,
Ⅰ. INTRODUCTION particularly deep learning, have enabled the creation of
In the realm of wind energy, ultra-short-term wind power complex models to surmount these obstacles[5]. Notably,
forecasting plays a vital role in the economic and reliable Convolutional Neural Networks (CNNs) and Recurrent Neural
operation of power systems[1]. With the escalating proliferation Networks (RNNs) are adept at processing time-series
information and extracting temporal and spatial characteristics.
Additionally, ensemble learning techniques such as Random CNN with dilated convolutions, the network gains proficiency
Forest and Gradient Boosting are extensively implemented to in time series feature extraction.
bolster the stability and precision of wind power predictions[6].
This paper seeks to refine traditional models by
reclassifying the dataset to minimize the impact of wind
variability and by optimizing ultra-short-term wind power
prediction through task-specific adjustments in separate
channels, aiming to substantially enhance prediction accuracy.
Ⅱ. RELATED WORK
To enhance the reliability and efficiency of wind power
integration into energy grids, recent advancements in ultra-
short-term wind power prediction have been particularly
significant. These developments leverage sophisticated
algorithms and machine learning techniques to achieve high Fig.1. Convolutional Neural Network
prediction accuracy. B. Structure of the LSTM model
For instance, An et al. [7] introduce a model that combines Long Short-Term Memory (LSTM) networks, a specialized
Phase Variation Mode Decomposition, Error Statistics subclass of recurrent neural networks, are particularly designed
Management Algorithm, and Deep Extreme Learning Machine to process and predict events in time series data characterized
to address the variability in wind power data effectively. by long intervals and delays [11], as illustrated in Figure 2.
Similarly, Yang, Chen, and Huang [8] employ fractal scaling These networks incorporate unique memory cell units that
factor transformations to improve multi-step predictions for enable them to maintain information over extended periods
large-scale wind farms, demonstrating the ability to capture without the risk of gradient vanishing or exploding, common
complex patterns in wind behavior. issues in traditional recurrent neural networks.
Moreover, Zhang et al. [9] use an LSTM network enhanced The core of an LSTM unit includes three gates: the forget
by an error-following forget gate to minimize the adverse gate, input gate, and output gate. The forget gate determines
effects of prediction errors, showcasing the effectiveness of which parts of the cell state are to be discarded and is defined
neural networks in this domain. The work by another team led by:
by An [10] applies the Sparrow Search Algorithm to optimize
deep learning models, further illustrating the shift from 𝑓! = 𝜎$𝑊" ⋅ [ℎ!#$ , 𝑥! ] + 𝑏" . (1)
traditional methods to more dynamic and adaptive approaches.
The input gate decides which new information is added to
These contributions collectively reflect the trend towards the cell state, and is given by:
employing more complex and accurate models to manage the
inherent unpredictability of wind power, underlining the critical 𝑖! = 𝜎(𝑊" ⋅ [ℎ!#$ , 𝑥! ] + 𝑏" ) (2)
role of advanced computational techniques in the renewable
Accompanying the input gate is the cell candidate, which
energy sector [11].
helps update the cell state with new candidate values, calculated
Ⅲ.DESIGNING A MULTI―CHANNEL FEATURE EXTRA as:
CTION MODEL
𝐶1! = tanh(𝑊% ⋅ [ℎ!#$ , 𝑥! ] + 𝑏% ) (3)
A. Convolutional Neural Network
Finally, the output gate controls what next-step information
A Convolutional Neural Network (CNN) primarily consists is output from the LSTM cell and is expressed as:
of a convolutional layer, a pooling layer, and a fully connected
layer, as depicted in Figure 1. The convolutional layer, the 𝑜! = 𝜎(𝑊& ⋅ [ℎ!#$ , 𝑥! ] + 𝑏& ) (4)
cornerstone of a CNN, employs convolutional kernels to
perform operations on the data, extracting its intrinsic features. ℎ! = 𝑜! ∗ tanh(𝐶! ) (5)
Subsequently, the pooling layer condenses the convolutional These mechanisms collectively allow LSTMs to effectively
layer's output to compress the data, thus enhancing the capture temporal correlations, enabling their application in
network's ability to generalize. CNNs' extensive application and environments where data dependencies extend over longer
outstanding performance in image, video, and speech periods.
processing underscore their significance as a crucial aspect of
deep learning. Their widespread applicability stems largely In the realm of deep learning, convolutional neural networks
from their localized receptive fields and shared weight (CNNs) excel in addressing classification challenges but fall
architecture[12]. In power prediction, the data comprises short in managing temporal dependencies in data. This paper
single-point time series with inherent seasonal patterns and leverages the robust memory capabilities of LSTMs to extract
human influences. The purely local receptive field of traditional time-dependent features from the first channel of actual data,
CNNs may overlook certain features. Consequently, there is a demonstrating their superior capacity to handle long-term
pressing need for networks capable of capturing the dependencies.
characteristics of time series data. By integrating the original
 Fig.2. LSTM Unit Structure
C. Ultra-short-term Power Prediction Feature Fig.3. Multi-channel Feature Extraction Model
Ultra-short-term power prediction enables the forecasting of Ⅳ. EXPERIMENTAL ANALYSIS AND DISCUSSION
a wind farm's energy output over the next four hours at 15-
minute intervals, utilizing the current conditions as a reference A.Data Preprocessing
point. This forecasting primarily relies on historical The historical data contains various anomalies, including
measurements and Numerical Weather Prediction (NWP) data. artificial power constraints, equipment maintenance, and
Historical data encompass time series information, such as past failures, as well as instances of high wind speeds coupled with
wind behavior, which can be approximately projected into the low power outputs. These anomalies disrupt the model's
future using the LSTM network's capability to remember training process by compromising its smoothness. To address
sequences[13]. NWP data provide comprehensive atmospheric this issue in our experiment, we employ the cross-validation
dynamics, illustrating the evolution of wind speed. By technique alongside the $3 \sigma$ rule to identify and rectify
analyzing various wind speed combinations, we can extract a these irregularities. The process begins by calculating the mean
broader range of features. and variance of the training data, as outlined in Equation6―8.
D. Multi-channel Feature Extraction Model )
1
In this study, we enhance the traditional artificial neural 𝑥'," = = 𝑥'," (6)
network by transitioning from a single-channel to a multi- 𝑁
'*$
channel structure, thereby minimizing cross-data interference )
and extracting more valuable information[14]. Figure depicts 1 +
the methodology, which involves creating three distinct dataset 𝜎"+ = =?𝑥'," − 𝑥'," A (7)
𝑁
combinations during the feature combination phase: '*$

(1) Historical time series data is processed using an LSTM C𝑥'," − 𝑥'," C > 3𝜎" 𝜀 (8)
network to distill time-dependent features.
Where 𝑥!,# represents the 𝑖 th data point on the 𝑛th day, with
(2) The addition of Numerical Weather Prediction (NWP)
𝑖 = 1,2,3, … ,96; 𝑥!,# denotes the sample mean; 𝜎#$ is the sample
information to corresponding time points in the measured data
variance; and 𝜀 serves as the threshold for anomaly detection.
enables the extraction of the interconnection between these data
Data points that fulfill the conditions of Equation 8 are
points and the NWP using a one-dimensional convolutional classified as outliers and must be excluded.
neural network.
(3) The integration of the complete dataset with a 1D B. Evaluation indicators
convolutional neural network serves to refine features further, To evaluate the prediction performance of the model, the root
augmenting those previously identified. Subsequently, the mean square error (RMSE) 𝐸 and the harmonic mean (HM)
feature fusion segment entails the amalgamation of the three accuracy 𝐹 are used as evaluation indices in this experiment.
sets of extracted features, which are then incorporated into the The RMSE accuracy is expressed as:
Fully Connected Layer (FC) for enhanced power forecasting
- $
1 𝑃&) − 𝑃*)
𝐸 = 1 − 𝐸%&'( = 1 − 23 5 6 9 : (9)
𝑁 𝐶+,
)./

Where: 𝐸%&'( is the root mean square error; 𝑘 is the

prediction point; 𝑁 is the number of points involved in the
calculation; 𝑃&) is the predicted value of the available powerat
the 𝑘 th point; 𝑃*) is the actual power at the 𝑘 th point; 𝐶ap
is the installed capacity of the wind farm. installed capacity of
the wind farm. The average accuracy of the reconciliation is [3] Habibi, M. R., Golestan, S., Guerrero, J. M., & Vasquez, J.
expressed as: C. (2023). Deep Learning for Forecasting-Based Applications
! in Cyber–Physical Microgrids: Recent Advances and Future
𝑃#0 |𝑃#0 − 𝑃#! | Directions. Electronics, 12(7), 1685.
𝐹 = 1 − 2 × 5 6@ 0 ! − 0.5@ × ! 9 (10)
𝑃# + 𝑃# ∑#./G𝑃#0 − 𝑃#! G
#./ [4] Khan, J., Lee, E., Balobaid, A. S., & Kim, K. (2023). A
C.Analysis of results comprehensive review of conventional, machine leaning, and
Wind farms A and B, each with an installed capacity of deep learning models for groundwater level (GWL)
99MW, have been selected for study. These farms are situated forecasting. Applied Sciences, 13(4), 2743.
in regions characterized by temperate and subtropical monsoon [5] Shabbir, N., Kütt, L., Jawad, M., Husev, O., Rehman, A. U.,
climates, respectively, and exhibit four distinct seasons: a mild
Gardezi, A. A., ... & Choi, J. G. (2022). Short-term wind energy
spring, a rainy summer, a crisp autumn, and a cold winter, all of
forecasting using deep learning-based predictive
which underscore the area's pronounced monsoon climate.
Table 1 presents a comparison of forecast results obtained from analytics. Comput. Mater. Contin, 72(1), 1017-1033.
various algorithms applied to the two fields' datasets. The [6] Huang, H. (2023). Ultra-short-term multi-step wind speed
results demonstrate that, in certain periods, the algorithm can prediction for wind farms based on adaptive noise reduction
effectively mitigate the phenomenon of ultra-short-term
technology and temporal convolutional network. arXiv preprint
forecast delays, thereby confirming its efficacy.
arXiv:2311.16198.
Table 1 Testset results
Error Accuracy [7] An, G., Chen, L., Tan, J., Jiang, Z., Li, Z., & Sun, H. (2022).
Ultra-short-term wind power prediction based on PVMD-
Wind farms Single-channel RMS Mulit-channel RMS ESMA-DELM. Energy Reports, 8, 8574-8588.
A 0.8312 0.85535 [8]Yang, M., Chen, X., & Huang, B. (2018). Ultra-short-term
B 0.8103 0.8182 multi-step wind power prediction based on fractal scaling factor
transformation. Journal of Renewable and Sustainable
Energy, 10(5).
V.CONCLUSIONS [9] Zhang, P., Li, C., Peng, C., & Tian, J. (2020). Ultra-short-
Comparative analysis of actual wind farm data demonstrates term prediction of wind power based on error following forget
that a multi-channel feature combination model for ultra-short- gate-based long short-term memory. Energies, 13(20), 5400.
term wind power prediction, based on a convolutional neural
network, outperforms traditional single-channel algorithms in [10] An, G., Jiang, Z., Chen, L., Cao, X., Li, Z., Zhao, Y., & Sun,
terms of its predictive accuracy and its resilience to noise within H. (2021). Ultra short-term wind power forecasting based on
the original dataset. sparrow search algorithm optimization deep extreme learning
machine. Sustainability, 13(18), 10453.
The design of the enhanced multi-channel neural network
creates three distinct channels to process historical measured [11] Sawant, M., Patil, R., Shikhare, T., Nagle, S., Chavan, S.,
data, numerical weather prediction data, and their interrelated Negi, S., & Bokde, N. D. (2022). A selective review on recent
couplings. Utilizing different networks for feature extraction advancements in long, short and ultra-short-term wind power
allows for the identification of a broader array of data prediction. Energies, 15(21), 8107.
characteristics, augments the capacity of a single network to
extract features, and mitigates the influence of variability in [12] Xiang, L., Liu, J., Yang, X., Hu, A., & Su, H. (2022). Ultra-
wind patterns, culminating in improved accuracy for ultra- short term wind power prediction applying a novel model
short-term power forecasting. named SATCN-LSTM. Energy Conversion and
REFERENCES Management, 252, 115036.
[13] Li, Y. L., Zhu, Z. A., Chang, Y. K., & Chiang, C. K. (2020,
[1] Chang, C., Meng, Y., Huo, J., Xu, J., & Xie, T. (2024). Ultra-
November). Short-term wind power forecasting by advanced
short-term wind power forecasting based on feature weight
analysis and cluster dynamic division. Journal of Renewable machine learning models. In 2020 International Symposium on
and Sustainable Energy, 16(2). Computer, Consumer and Control (IS3C) (pp. 412-415). IEEE.
[14] Khan, M. A., Khan, I. A., Shah, S., Mohammed, E. A., &
[2] Sharma, V. (2018). Deterministic and Probabilistic
Jadoon, W. (2024). Short-term wind power forecasting through
Forecasting for Wind and Solar Power Using Advance Data
Analytics and Machine Learning Techniques (Doctoral stacked and bi directional LSTM techniques. PeerJ Computer
dissertation, The University of North Carolina at Charlotte). Science, 10, e1949.