Low-dimensional Models in Spatio-Temporal Wind Speed Forecasting

by   Borhan M. Sanandaji, et al.

Integrating wind power into the grid is challenging because of its random nature. Integration is facilitated with accurate short-term forecasts of wind power. The paper presents a spatio-temporal wind speed forecasting algorithm that incorporates the time series data of a target station and data of surrounding stations. Inspired by Compressive Sensing (CS) and structured-sparse recovery algorithms, we claim that there usually exists an intrinsic low-dimensional structure governing a large collection of stations that should be exploited. We cast the forecasting problem as recovery of a block-sparse signal x from a set of linear equations b = Ax for which we propose novel structure-sparse recovery algorithms. Results of a case study in the east coast show that the proposed Compressive Spatio-Temporal Wind Speed Forecasting (CST-WSF) algorithm significantly improves the short-term forecasts compared to a set of widely-used benchmark models.



There are no comments yet.


page 3


On cointegration for modeling and forecasting wind power production

This study evaluates the performance of cointegrated vector autoregressi...

Wind models and cross-site interpolation for the refugee reception islands in Greece

In this study, the wind data series from five locations in Aegean Sea is...

Deep Spatio-Temporal Wind Power Forecasting

Wind power forecasting has drawn increasing attention among researchers ...

Short-Term Wind-Speed Forecasting Using Kernel Spectral Hidden Markov Models

In machine learning, a nonparametric forecasting algorithm for time seri...

Performance Analysis of ANFIS in short term Wind Speed Prediction

Results are presented on the performance of Adaptive Neuro-Fuzzy Inferen...

Forecasting wind power - Modeling periodic and non-linear effects under conditional heteroscedasticity

In this article we present an approach that enables joint wind speed and...

Adaptive Generalized Logit-Normal Distributions for Wind Power Short-Term Forecasting

There is increasing interest in very short-term and higher-resolution wi...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

I-a Variable Energy Resources

Many countries in the world as well as many states in the U.S. have mandated aggressive Renewable Portfolio Standards. Among different renewable energy resources, wind energy itself is expected to grow to provide between to of the world’s global electricity by 2050. According to another study, the world total wind power capacity has doubled every three years since 2000, reaching an installed capacity of 197 GW in 2010 and 369 GW in 2014 [1, 2]. The random nature of wind, however, makes it difficult to achieve the power balance needed for its integration into the grid [3, 4]. The use of ancillary services such as frequency regulation and load following to compensate for such imbalances[5, 6, 7, 8] is facilitated by accurate forecasts [9, 10].

I-B Wind Energy Forecasting Methods

One can directly attempt to forecast wind power. An alternative approach is to forecast the wind speed and then convert it to wind power using given power curves. This approach will accommodate different wind turbines installed in a wind farm experiencing the same wind speed but resulting in different wind power generation. We focus on wind speed forecasting in this paper. Wind speed forecasting methods can be categorized to different groups: (i) model-based methods such as Numerical Weather Prediction (NWP) vs. data-driven methods, (ii) point forecasting vs. probabilistic forecasting, and (iii) short-term forecasting vs. long-term forecasting. This paper is concerned with short-term point forecasting using both temporal data as well as spatial information. For a more complete survey of wind speed forecasting methods see [11] and [12] among others.


I-C Spatio-Temporal Wind Speed Forecasting

There is a growing interest in the so-called spatio-temporal forecasting methods that use information from neighboring stations to improve the forecasts of a target station, since there is a significant cross-correlation between the time series data of a target station and its surrounding stations. We review some of the spatio-temporal forecasting methods. Gneiting et al. [13] introduced the Regime Switching Space-Time Diurnal (RSTD) model for average wind speed data based on both spatial and temporal information. This method was later improved by Hering and Genton [14] who incorporated wind direction in the forecasting process by introducing Trigonometric Direction Diurnal (TDD) model. Xie et al. [15] also considered probabilistic TDD forecast for power system economic dispatch. Dowell et al. [16] employed a multi-channel adaptive filter to predict the wind speed and direction by taking advantages of spatial correlations at numerous geographical sites. He et al. [17]

presented Markov chain-based stochastic models for predictions of wind power generation after characterizing the statistical distribution of aggregate power with a graph learning-based spatio-temporal analysis. Regime-switching models based on wind direction are studied by Tastu et al. 

[18] where they consider various statistical models, such as ARX models, to understand the effects of different variables on forecast error characteristics. A methodology with probabilistic wind power forecasts in the form of predictive densities taking the spatial information into account was developed in [19]. Sparse Gaussian Conditional Random Fields (CRFs) have also been deployed for probabilistic wind power forecasting [20]. See [21] for a comprehensive review of the state-of-the-art methods.

I-D Our Contribution

Inspired by Compressive Sensing (CS) and structured-sparse recovery algorithms, we claim that there usually exists an intrinsic low-dimensional structure governing the interactions among a large collection of weather stations. Such low-dimensional models should be exploited in the forecasting process. To this end, we cast the forecasting problem as the recovery of a block-sparse signal from a set of linear equations for which we propose novel structure-sparse recovery algorithms. As a case study, we apply our proposed forecasting algorithm to the data recorded from 57 measuring locations (a combination of airports and weather stations) in a region in the east coast including Massachusetts, Connecticut, New York, and New Hampshire. The results lead to a considerable improvement of the short-term forecasts compared to a set of widely-used benchmark models and advanced spatio-temporal approaches.

I-E Paper Organization

In Section II, we formulate the forecasting problem. The proposed forecasting algorithms and related concepts are presented in Section III. We apply the proposed methods to real wind speed data and compare the results with other benchmark methods in Section IV. Section V presents our concluding remarks and possible future directions.

Ii Multivariate Autoregressive (M-AR) Model

Autoregressive (AR) models assume that the output variable of a system can be well presented as a weighted linear combination of its own previous values.  Multivariate Autoregressive (M-AR

) (a.k.a., Vector Autoregressive) models generalize this approach to multivariate time series. Let

be a -dimensional output measurement (e.g., wind speeds at weather stations) at time . An M-AR model of order is written as


where is a coefficient matrix associated with the -th time lag and is a Gaussian noise. Using a different notation, let be the wind speed of the -th station at sample time for . For each station, the M-AR model (2) can be re-written in a matrix-vector product format as in (1), where . In the training stage, the goal is to find a coefficient vector that best explains the observations and . As seen from (1), has a block structure as the coefficients corresponding to each station appear in one vector-block.

Iii Compressive Spatio-Temporal
Wind Speed Forecasting (CST-WSF)

We believe that among a large collection of stations, only a few of them have a strong correlation with the target station. We show that under the assumption of sparsity of the interconnections (that is, assuming only a few weather stations contribute to the output of the target station), there will be a distinct structure to the solution that we are seeking. In particular, a typical coefficient vector under our model assumptions will have very few non-zero entries, and these non-zero entries will be clustered in few locations. Vectors with such structure are called block-sparse. The number of blocks corresponds to the number of links that contribute to the output of the target station. For a given target station, we then solve the minimization problem:


We call this approach  Compressive Spatio-Temporal Wind Speed Forecasting (CST-WSF), as it is inspired by Compressive Sensing (CS) and structured-sparse recovery.

Iii-a Background on Cs

CS enables recovery of an unknown signal from its underdetermined set of measurements under the assumption of sparsity of the signal and under certain conditions on the measurement matrix  [22]. The CS recovery problem can be viewed as recovery of a -sparse signal from its observations where is the measurement matrix with (in many cases ). A -sparse signal is a signal of length with non-zero entries where . Since the null space of is non-trivial, there are infinitely many candidate solutions to the equation ; however, CS recovery algorithms exploit the fact that, under certain conditions on , only one candidate solution is suitably sparse. An interested reader can refer to [23, 24] for several proposed recovery conditions.

Iii-B Uniform Cst-Wsf

We adapt our CST-WSF algorithm from tools proposed in the CS literature for recovery of a block-sparse signal .

Definition 1 (Block -Sparse Signal)

Let be a concatenation of vector-blocks , i.e.,

where . A signal is called block -sparse if it has non-zero blocks.

Several extensions of the standard CS recovery algorithms can account for additional structure in the sparse signal to be recovered [25, 26]. Among these, the Block Orthogonal Matching Pursuit (BOMP) algorithm [26] is designed to exploit block sparsity due to its flexibility in recovering block-sparse signals of different sparsity levels and its low computation complexity [27]. In a more general setting, BOMP has been recently used for topology identification of interconnected dynamical systems (e.g., see [28]).


Iii-C Nonuniform Cst-Wsf

In a uniform CST-WSF, the assumption is that a uniform M-AR model as given in (1) is governing the interactions between stations. In other words, we assume that the target station and its surrounding stations are related by AR models of the same order. In this section, we consider a more generalized version of the CST-WSF algorithm where the target station and its surrounding stations are related by AR models of different orders. This model structure, called Nonuniform Multivariate Autoregressive (NM-AR) distinguishes between the stations with high and low cross-correlation with the target station. Let be the order associated with the -th station for . An NM-AR version of (1) can be considered as given in (4), where and . This results in a nonuniform block-sparse coefficient vector whose blocks have different length.

Definition 2 (Nonuniform Block -Sparse Signal)

Let as a concatenation of vector-blocks where . A signal is called nonuniform block -sparse if it has non-zero blocks.

Remark 1

One should note that our definition of nonuniform block -sparse signals is a generalization of the conventional definition of block -sparse signals (Definition 1) where all blocks have the same length, i.e., .

Given , the BOMP algorithm can be used for recovery of with . In order to find the set of order , we use a correlation analysis. We then adjust the orders to achieve the best prediction performance.

Iv Case Study of 57 Stations in East Coast

We apply our proposed CST-WSF algorithms to real wind speed data. East coast states are good candidates for our study because: (i) wind speed profiles are higher and (ii) there are more stations in a close vicinity in these states.

Iv-a Data Description

We use hourly wind speed data from Meteorological Terminal Aviation Routine (METAR) weather reports of 57 stations in east coast including Massachusetts, Connecticut, New York, and New Hampshire [29]. Fig. 1 depicts the area under study and the location of these 57 stations. The target station  Nantucket Memorial Airport (ACK) (circled in red) is located on an island and is subject to wind profiles with high ramps and speeds due to the fact that the surrounding surface has very low roughness heights. Furthermore, this area has good correlations with other stations owing to the fact the prevailing wind direction of this region is mainly northwest or southeast.

Fig. 1: Map of the area under study. The measuring locations in east coast are shown with yellow points. Circled in red is the target station ACK.

A time period from January 6, 2014 to February 20, 2014 is considered in our simulations. This time period has the most unsteady wind conditions throughout the year. The data is divided to two subsets: (i) training subset from January 6, 2014 to February 6, 2014 (a period of 30 days) and (ii) validation subset from February 6, 2014 to February 20, 2014 (a period of 14 days).

Iv-B Comparison with Other Benchmark Algorithms

In order to better gauge the effectiveness of the proposed algorithm, we compare CST-WSF with other proposed benchmark algorithms in temporal and spatio-temporal wind speed forecasting. For temporal wind speed forecasting, we first consider persistence forecasting method which simply uses the last measured value for the forecast interval. Any algorithm that can improve upon persistence forecasting is judged to be an effective algorithm. We also consider AR models as well as an advanced prediction model that combines Wavelet Transform (WT) with

Artificial Neural Network

(ANN). The latter method is shown to have the capability of capturing nonlinearity in the wind speed time series. In this model, briefly, the volatile wind speed series is first cut up by the WT

into a number of better-behaved sub-series in various frequency bands. Subsequently, estimates of each extracted sub-series are carried out separately employing the

ANN mode. Speed predictions are then reconstructed except for the highest frequency band which represents the most fluctuating part of the wind speed series (see [30] for more details). Multi-step predictions were performed in a recursive manner for a period of 14 consecutive days with 6 hour-ahead updates. The prediction results are given in Fig. 2. As can be seen, the considered temporal prediction methods provide reasonable forecasting compared to persistence model. ANN-based model outperforms the AR model.

(a) Persistence forecasting
(b) AR model of order 3
(c) WT-ANN model
Fig. 2: Comparison of different temporal forecasting algorithms.

We also consider two spatio-temporal forecasting methods. We first employed an advanced ANN-based spatio-temporal model [31]. We also employed a Least Squares (LS) M-AR spatio-temporal forecasting approach [18]. Fig. 3 depicts how incorporation of spatial information improves the forecasting performance as compared to temporal methods.

(a) Spatio-temporal ANN model
(b) Spatio-temporal LS M-AR model
Fig. 3: Comparison of different spatio-temporal forecasting algorithms.
(a) Uniform CST-WSF with equal orders of 3.
(b) Nonuniform CST-WSF with nonuniform orders.
Fig. 4: Comparison of the uniform and nonuniform CST-WSF.

Iv-C Uniform and Nonuniform Cst-Wsf

We now apply our proposed uniform and nonuniform CST-WSF methods. Note that in our simulations a new coefficient vector is obtained every 6 hours (equivalently, every 6 time steps as each time step is 1 hour). This is the considered prediction time horizon. Also, we follow a recursive approach in prediction. That is, the wind speed predictions at time for all stations () are included in the matrix for predicting the wind speed at time () and so on. This recursive process goes on for 6 time steps. The elements of are then completely updated with real measurements and the recursive process continues for another 6 time steps.

Figure 4(a) shows the result of the uniform CST-WSF. The result is superior to all benchmark approaches discussed in the previous section. We then apply the nonuniform CST-WSF algorithm. The result is illustrated in Fig. 4(b). In order to demonstrate the effectiveness of the proposed forecasting models, the associated Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) values, which are the most common performance metrics in the wind forecasting literature, are listed in Table I for all of the wind speed forecasting methods considered in this paper. The MAE provides the average deviation between the measured and predicted data while the RMSE gives higher weights to larger error values by squaring the differences. Moreover, the Normalized Root Mean Squared Error (NRMSE) which is the RMSE normalized by the range of observed data, is calculated to provide a scale-independent error measure. Evidently the proposed uniform and nonuniform CST-WSF methods outperform the other considered temporal and spatio-temporal methods. The best prediction is provided by the nonuniform CST-WSF. Considering the NRMSE as an example, nonuniform CST-WSF approach provides a reduction of 38%, 36.2% and 28.7% as compared to the considered temporal methods (persistence forecasting, AR of order 3, and WT-ANN models) and a reduction of 24% and 20% as compared to the considered spatio-temporal methods (ANN-based ST and LS-based ST), respectively.

Forecasting approach MAE (m/s) RMSE (m/s) NRMSE (%)
Persistence Forecasting 2.1441 2.8334 16.86
AR of order 1 2.0742 2.7629 16.44
AR of order 3 2.0696 2.7560 16.40
WT-ANN 1.8200 2.4671 14.68
ANN-based ST 1.7981 2.2997 13.69
LS-based ST 1.7234 2.1983 13.08
CST-WSF of order 3 1.5665 2.0595 12.25
Nonuniform CST-WSF 1.3442 1.7586 10.46
TABLE I: Statistical Error Measure Comparison of Different Methods
Fig. 5: Block-sparse coefficient vector. The red dashed lines specify 57 vector-blocks of the coefficient vector. (a) Uniform CST-WSF of order 3. (b) Nonuniform CST-WSF with different orders (vector-block lengths).

Fig. 5 illustrates the corresponding block-sparse coefficient vectors for the uniform and nonuniform CST-WSF methods at the training stage. As can be seen, only a few of the blocks in uniform and nonuniform CST-WSF are non-zero, resulting in a block-sparse . This block-sparse structure appears in all of the other calculated coefficient vectors (with different block-sparsity pattern) as we move over prediction horizon time and further confirms our motivation behind exploiting the intrinsic low-dimensional models in spatio-temporal wind speed forecasting. It is worth noting that the proposed CST-WSF methods have a much shorter computational time as compared to other ANN-based methods and the average computational time for other proposed short-term forecasting methods in the literature [15, 32]. For instance, the total simulation time of nonuniform CST-WSF approach is almost the half of the time required for the predictions with WT-ANN model and approximately 15% smaller than that of LS M-AR spatio-temporal model in this study.

V Conclusion

We proposed two spatio-temporal wind speed forecasting methods, called uniform and nonuniform CST-WSF. The methods are inspired by CS and structured-sparse recovery algorithms, where we claim that there usually exists an intrinsic low-dimensional structure governing a large collection of stations. The results of a case study show that the proposed approaches considerably improves the short-term forecasts compared to a set of widely-used benchmark models.

As future directions, we plan to apply the proposed CST-WSF to a much larger set of stations. Incorporating other variables (such as temperature, pressure, etc.) in the wind speed forecasting is another research path. Yet another direction is to investigate using sparsity-based ideas in probabilistic forecasting methods. Such information about the forecast errors are useful for Transmission System Operators (TSOs), Independent Power Producers (IPPs), etc., in evaluating the economic and technical risks due to uncertainty.


  • [1] International Energy Agency (IEA), “Technology roadmap: Wind energy,” 2013. [Online]. Available: http://www.iea.org/publications/freepublications/publication/Wind˙2013˙Roadmap.pdf
  • [2] California Energy Commission, “California renewable energy overview and programs,” 2013. [Online]. Available: http://www.energy.ca.gov/renewables/index.html
  • [3] J. C. Smith, M. R. Milligan, E. A. DeMeo, and B. Parsons, “Utility wind integration and operating impact state of the art,” IEEE Transactions on Power Systems, vol. 22, no. 3, pp. 900–908, 2007.
  • [4] E. Bitar, R. Rajagopal, P. Khargonekar, and K. Poolla, “The role of co-located storage for wind power producers in conventional electricity markets,” Proceedings of the American Control Conference, pp. 3886–3891, 2011.
  • [5] D. S. Callaway, “Tapping the energy storage potential in electric loads to deliver load following and regulation, with application to wind energy,” Energy Conversion and Management, vol. 50, no. 5, pp. 1389–1400, 2009.
  • [6] A. Tascikaraoglu, M. Uzunoglu, B. Vural, and O. Erdinc, “Power quality assessment of wind turbines and comparison with conventional legal regulations: A case study in turkey,” Applied Energy, vol. 88, no. 5, pp. 1864–1872, 2011.
  • [7] B. M. Sanandaji, H. Hao, and K. Poolla, “Fast regulation service provision via aggregation of thermostatically controlled loads,” Proceedings of the th Hawaii International Conference on System Sciences, pp. 2388–2397, 2014.
  • [8] J. Qin, Y. Chow, J. Yang, and R. Rajagopal, “Modeling and online control of generalized energy storage networks,” Proceedings of th International Conference on Future Energy Systems, pp. 27–38, 2014.
  • [9] J. Wang, M. Shahidehpour, and Z. Li, “Security-constrained unit commitment with volatile wind power generation,” IEEE Transactions on Power Systems, vol. 23, no. 3, pp. 1319–1327, 2008.
  • [10] A. Tascikaraoglu, O. Erdinc, M. Uzunoglu, and A. Karakas, “An adaptive load dispatching and forecasting strategy for a virtual power plant including renewable energy conversion units,” Applied Energy, vol. 119, pp. 445–453, 2014.
  • [11] X. Zhu and M. G. Genton, “Short-term wind speed forecasting for power system operations,” International Statistical Review, vol. 80, no. 1, pp. 2–23, 2012.
  • [12] A. Tascikaraoglu and M. Uzunoglu, “A review of combined approaches for prediction of short-term wind speed and power,” Renewable and Sustainable Energy Reviews, vol. 34, pp. 243–254, 2014.
  • [13] T. Gneiting, K. Larson, K. Westrick, M. G. Genton, and E. Aldrich, “Calibrated probabilistic forecasting at the stateline wind energy center: The regime-switching space–time method,” Journal of the American Statistical Association, vol. 101, no. 475, pp. 968–979, 2006.
  • [14] A. S. Hering and M. G. Genton, “Powering up with space-time wind forecasting,” Journal of the American Statistical Association, vol. 105, no. 489, pp. 92–104, 2010.
  • [15] L. Xie, Y. Gu, X. Zhu, and M. G. Genton, “Short-term spatio-temporal wind power forecast in robust look-ahead power system dispatch,” IEEE Transactions on Smart Grid, vol. 5, no. 1, pp. 511–520, 2014.
  • [16] J. Dowell, S. Weiss, D. Hill, and D. Infield, “Short-term spatio-temporal prediction of wind speed and direction,” Wind Energy, 2013.
  • [17] M. He, L. Yang, J. Zhang, and V. Vittal, “A spatio-temporal analysis approach for short-term forecast of wind farm generation,” IEEE Transactions on Power Systems, vol. 29, no. 4, pp. 1611–1622, 2014.
  • [18] J. Tastu, P. Pinson, E. Kotwa, H. Madsen, and H. A. Nielsen, “Spatio-temporal analysis and modeling of short-term wind power forecast errors,” Wind Energy, vol. 14, no. 1, pp. 43–60, 2011.
  • [19] J. Tastu, P. Pinson, P.-J. Trombe, and H. Madsen, “Probabilistic forecasts of wind power generation accounting for geographically dispersed information,” IEEE Transactions on Smart Grid, vol. 5, no. 1, pp. 480–489, 2014.
  • [20] M. Wytock and Z. Kolter, “Large-scale probabilistic forecasting in energy systems using sparse gaussian conditional random fields,” Proceedings of the nd IEEE Conference on Decision and Control, pp. 1019–1024, 2013.
  • [21] Y. Zhang, J. Wang, and X. Wang, “Review on probabilistic forecasting of wind power generation,” Renewable and Sustainable Energy Reviews, vol. 32, pp. 255–270, 2014.
  • [22] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,” IEEE Transactions on Information Theory, vol. 52, no. 2, pp. 489–509, 2006.
  • [23] J. Tropp, “Greed is good: Algorithmic results for sparse approximation,” IEEE Transactions on Information Theory, vol. 50, no. 10, pp. 2231–2242, 2004.
  • [24] D. Donoho and X. Huo, “Uncertainty principles and ideal atomic decomposition,” IEEE Transactions on Information Theory, vol. 47, no. 7, pp. 2845–2862, 2001.
  • [25] R. G. Baraniuk, V. Cevher, M. F. Duarte, and C. Hegde, “Model-based compressive sensing,” IEEE Transactions on Information Theory, vol. 56, no. 4, pp. 1982–2001, 2010.
  • [26] Y. C. Eldar, P. Kuppinger, and H. Bölcskei, “Block-sparse signals: uncertainty relations and efficient recovery,” IEEE Transactions on Signal Processing, vol. 58, no. 6, pp. 3042–3054, 2010.
  • [27] B. M. Sanandaji, T. L. Vincent, and M. B. Wakin, “A review on sufficient conditions for structure identification of interconnected systems,” Proceedings of the th IFAC Symposium on System Identification, pp. 1623–1628, 2012.
  • [28] B. M. Sanandaji, “Compressive system identification (CSI): Theory and applications of exploiting sparsity in the analysis of high-dimensional dynamical systems,” Ph.D. dissertation, Colorado School of Mines, 2012.
  • [29] Iowa Environmental Mesonet, “ASOS historical data,” 2014. [Online]. Available: http://mesonet.agron.iastate.edu/ASOS/
  • [30] A. Tascikaraoglu, M. Uzunoglu, and B. Vural, “The assessment of the contribution of short-term wind power predictions to the efficiency of stand-alone hybrid systems,” Applied Energy, vol. 94, pp. 156–165, 2012.
  • [31] M. Bilgili, B. Sahin, and A. Yasar, “Application of artificial neural networks for the wind speed prediction of target station using reference stations data,” Renewable Energy, vol. 32, no. 14, pp. 2350–2360, 2007.
  • [32] A. Tascikaraoglu, A. Boynuegri, and M. Uzunoglu, “A demand side management strategy based on forecasting of residential renewable sources: A smart home system in turkey,” Energy and Buildings, vol. 80, pp. 309–320, 2014.