Homogenization of Tianjin monthly near-surface wind speed using RHtestsV4 for 1951–2014

3.1 Creation of the reference series

A much stronger autocorrelation commonly occurs in climate time series on a monthly scale, and is easily affected by local weather disturbances and seasonal variability. Accordingly, to obtain the best results, here the monthly candidate and its reference series in the calendar order were transformed into series in a calendar year with the same month; this technique has not been used in previous studies on the homogenization of near-surface wind speed (Wan et al. 2010; Li et al. 2011; Cusack 2013; Minola et al. 2016; Azorin-Molina et al. 2014, 2016). The change signal in the climate time series may be a systematic error, but it may also be a periodic variation in local climate (Li 2016). The rapid urbanization in Tianjin is remarkable, and has likely impacted regional climate and environment patterns, such as climate warming, air pollution, and anthropogenic changes. Therefore, the creation of a reference series in this study is still motivated by the same factors as previous historical homogenization data products in Tianjin (Si and Xie 2015a; Si and Xu 2015b), to preserve the objectivity of climate data. Therefore, the method uses observations from nearby stations in the Tianjin network (Fig. 1), instead of using the geostrophic wind speed series reflecting atmospheric circulation (Wan et al. 2010; Minola et al. 2016) or model simulation data not affected by human activities on the ground (Azorin-Molina et al. 2014).

3.2 Determining discontinuities in the wind speed series

Good metadata are one of the most indispensable elements for homogenizing the observed data (Aguilar et al. 2003). Here, the judgment of breakpoints in wind speed series was mainly focused on detailed station metadata information, taking their accuracy and objectivity into account from a subjective perspective. This is different from the processes used in Azorin-Molina et al. (2014), who had limited metadata, where breakpoints were fixed only using a mathematical approach based on statistical significance. Therefore, in this study, station metadata from NMIC and a large number of historical archive information about each station obtained from the Tianjin Meteorological Archives were used, providing a basis for determining and correcting breakpoints in the wind speed series. The detected series Z _i from 13 stations on an annual or monthly scale was tested using PMT (Wang et al. 2007), and potential breakpoint locations were fixed using station metadata. Combined with the test results from creating the corresponding reference series, significant breakpoints were clearly supported by station metadata and removed from the detected annual and monthly series at the same time as adjustment. The statistically significant confidence level is at 95%.

4.1 Statistics on the number of breakpoints and their related reasons

Figure 2 shows the number of breakpoints in the period from 2002 to 2004 as highest in monthly mean surface wind speed series covering 1951–2014; the metadata indicates this was primarily due to the effects of automation and relocation. The number of breakpoints in the 1969–1970 and 1988–1989 periods was primarily caused by instrument change and relocation, also referring to the metadata.

Open image in new window

4.2 Causes of the inhomogeneities in the wind speed series

Figure 3 shows the time series of annual mean raw and adjusted monthly mean wind speed at four representative stations. The annual wind speed series values at station 54525 (Fig. 3a) were abnormally larger before the year 1970, and for stations 54526 and 54623 with the shift before year 1974 and 1995, respectively, but the values at station 54529 (Fig. 3b) are unusually smaller before the shift in 1997 due to the instrument height increase. However, the shifts in the wind speed series at each station caused by instrument change were almost entirely corrected with homogenization, which makes the series relatively flat and continuous, shown in Fig. 3a, b.

Open image in new window

The shifts in monthly mean surface wind speed series at Tianjin (54527) and Hangu (54530) stations were both due to the effects of automation, based on their metadata. These shifts occurred between 2002 and 2004, especially for station 54527, where the monthly series from January to December in 2003 all have breakpoints. The annual values of wind speed based on the raw monthly series at station 54527 (Fig. 3c) were abnormally smaller before 2004 due to automation, but much larger for station 54530 prior to 2004. Similarly, data homogenization almost corrected the shifts at the two stations caused by automation, making their wind speed data relatively reasonable.

Relocation had a much greater effect on station Dagang (54645) than on station Beichen (54528), which were both significantly impacted. The shifts in station 54645 due to relocation were found in the other 11 months except February 2003, resulting in abnormally larger wind speed values before 2003 (Fig. 3d). The metadata information indicated that automation resulted in an instrument-type change, from EL anemometrographs to 15-1A-type EL wind speed sensors. Moreover, the effects of relocation impacted two factors: the observation environment and instrument height (Wan et al. 2010). Statistics in Table 3 also show that shifts in wind speed series at the City (54517) and Jinghai (54619) stations were both due to the effects of instrument change combined with automation, but the shift amplitude was lower.

Accordingly, the predominant reason for inhomogeneities in the Tianjin monthly mean surface wind speed series was instrument change, including changes in type and height. The automation effect increased the impact of instrument change. However, homogenization seemed to correct the shifts in the time series caused by non-climate factors at each station, making the wind speed characteristics relatively reasonable. This is consistent with many previous studies on the homogenization of near-surface wind speed in China (Liu 2000; Cao et al. 2010).

4.3 Probability density distribution of all QM adjustments

Figure 4 shows the probability density distribution of all QM adjustments applied to monthly mean surface wind speed series identified as inhomogeneity at the 10 Tianjin stations. The curve of empirical estimate indicates that the positive adjustments were larger than those of negative, 55.4 and 44.6% in the whole QM adjustments, respectively. The amplitudes of positive adjustments of probability density above 0.2 ranged from 0.2 to 1.2 m s⁻¹, 85% of the positive QM adjustments derived from the Kernel density estimate. The negative adjustments ranged from −0.1 to −1.2 m s⁻¹, 80% of the negative QM adjustments. The monthly mean surface wind speed series at the 10 inhomogeneous stations range had QM adjustments from −3.4 to 3.2 m s⁻¹, with mean and median values of 0.03 and 0.3 m s⁻¹, respectively.

Open image in new window

5.1 Variance comparison before and after adjustment

The variance is used to measure the average data change around its mean in the sample. In general, amplitude differences between climate variables in different areas can also be judged using variance (Huang 2000). Figure 5 provides the variance distribution of the annual time series based on the raw and adjusted monthly mean surface wind speed. Figure 5a, b indicates that annual variances based on the adjusted monthly data were less than those calculated from the raw ones, and the average variances for the entire Tianjin area were 0.343 and 0.416 m s⁻¹, respectively. These characteristics were also reflected in the seasonal wind speed time series at each station (figure omitted). Differences in the variance between raw and adjusted data from Fig. 6 indicate that station Dagang (54645) was the station with the largest difference in annual variance based on raw and adjusted monthly wind speed data, with a value of 0.75 m s⁻¹, especially in spring, decreasing to about 1.18 m s⁻¹. The metadata information suggests that station 54645 was influenced more by growing urbanization since 2003 due to relocation from rural to suburban, resulting in a smaller mean wind speed (Zhou and Shu 1994; Cui and Shi 2012; Wu et al. 2016). However, homogenization corrected the bias caused by relocation, making the wind speed change more reasonable for station 54645. Accordingly, this implies that the homogeneity adjustment reduced the large dispersion in mean wind speed data caused by some inhomogeneity factors, improving wind speed change stability. In contrast, annual and seasonal variances based on the adjusted monthly data were larger than those calculated using the raw data from station Tianjin (54527); the annual variance increased by 0.27 m s⁻¹ compared with the raw data, and similarly, the amount increased in spring with a maximum of 0.41 m s⁻¹. From the metadata, the location of station 54527 was in low-lying land surrounded by building obstacles before 2010; but this effect was largely a local environmental impact, rather than urbanization. The surrounding area was more flat and open after 2010; thus, the observed mean wind speed increased. However, the reason for the shifts in the wind speed series at station 54527 in Table 3 was automation (also may be called as instrument-type change); the effect of relocation seems to be concealed. The analysis results from Sect. 4.2 indicate that the relocation effect was also a type of instrument change. Therefore, the instability in wind speed changes after adjustment at station 54527 is reasonable. Moreover, Fig. 6 also shows that there is a clear increase in variance in the winter wind speed series at Tanggu station (54623) after adjustment. Station 54623 is the only land observation located near the coast in Tianjin; therefore, a winter variance change in wind speed series due to weather systems is reasonable; winter brings frequent cold air activity, followed by more sea gales than in other seasons. Similar findings were found by William and Martha (2014) who researched the Beaufort/Chukchi Sea coastal region. Coastal locations exhibited higher wind speeds than interior stations in all months of the year, with maximum monthly wind speeds occurring in the winter months. Minola et al. (2016) found similar results.

Open image in new window

Open image in new window

5.2 Spatial distribution of trends comparison before and after adjustment

Figure 7 shows the spatial distribution of trends in annual and seasonal time series based on the raw and adjusted monthly mean wind speed. The Dagang (54645) area shows decreasing trends in the annual mean wind speed (Fig. 7a) in Tianjin, which was transferred to the City area (54517) (Fig. 7b) with homogeneity adjustment. This is also reflected in the trend distribution in the seasonal mean wind speed (Fig. 7d, f, h, j), which is consistent with the climate change characteristics in Tianjin. Station metadata indicates that City station (54517) was impacted more prominently by urbanization due to its location in the center of Tianjin, and the decreasing trend in wind speed is clear compared to other stations (Hou et al. 2013; Dou et al. 2015). In addition, Ninghe (54529) also had a decreasing trend, displayed in Fig. 7b, compared to the raw wind speed trend distribution (Fig. 7a). Station 54529 was severely impacted by urbanization in the last few years; the surrounding environment did not conform to observation conditions required by the ground meteorological service, thereby station relocation has being carried out. This demonstrates that recent urbanization was not associated with continuous large-scale city construction; on the contrary, the surrounding open rural and suburban areas were the first targets for urban construction. As a result, trends in near-surface wind speed series at stations in open rural environment decreased due to building growth, which impacted stations from urbanization. This is also one of the reasons that there will be worse climate at rural stations in the next few years compared to urban stations (Si et al. 2014).

Open image in new window Open image in new window Open image in new window

The spatial distributions of seasonal mean wind speeds were more consistent when homogenized compared to raw trends (Fig. 7c, e, g, i). Trends in the northern and central regions were smaller, while in the southwest (city area) they had large decreasing values (Fig. 7d, f, h, j). The trend distributions were similar in spring and winter (Fig. 7d, j) with larger decreasing trends, and almost identical in summer and autumn (Fig. 7f, h) with smaller reducing trends. This is consistent with monsoon circulation dominating climate in the Tianjin area.

5.3 Trend amplitudes before and after adjustment

As shown in Fig. 8, the wind speed series after adjustment at all 10 stations were coincident, with a reducing trend at the 95% significant level. This result is consistent with current researches on assessment of observed near-surface wind speed trends in China (Rong and Liang 2008; Cheng 2010; Zou et al. 2010; Fu et al. 2011; Guo et al. 2011; Chen et al. 2013; Zhang et al. 2014; Xiong 2015), and also similar to trends reported for other parts of the world (Pryor et al. 2009; Wan et al. 2010; William and Martha 2014). Based on the adjusted data, the trends in the regional average annual and seasonal (spring, summer, autumn, and winter) wind speed series were −0.277, −0.325, −0.228, −0.223, and −0.336 m s⁻¹/10a, respectively (Fig. 8b), significantly less than the raw values of −0.335, −0.404, −0.271, −0.296, and −0.373 m s⁻¹/10a (Fig. 8a). This feature was also reflected in the spatial distribution of wind speed variation, shown in Fig. 7. The statistics based on the adjusted data here were generally consistent with those from Li et al. (2011), who evaluated the regional trends in annual and seasonal wind speed series based on homogenized observations in Beijing, −0.260, −0.300, −0.120, −0.220, and −0.390 m s⁻¹/10a, respectively (Table 4). Both this study and the prior study indicate that the decreasing amplitudes of wind speed trend in spring and winter were much larger than those in summer and autumn; this is also reflected in studies on annual and seasonal wind speed change over China (Xu et al. 2006; Jiang et al. 2010; Guo et al. 2011; Fu et al. 2011; and Chen et al. 2013). Accordingly, based on the adjusted data, characteristics of near-surface wind speed change in Tianjin are consistent with those of regional background climate in China.

Open image in new window

Table 4

Mean wind speed trend magnitudes over China from previously published studies, units: m s⁻¹/10a

Previous paper	Time series	Region	Annual	Spring	Summer	Autumn	Winter
Li et al. (2011)	1960–2008	Beijing	−0.260	−0.300	−0.120	−0.220	−0.390
Xu et al. (2006)	1969–2000*	North China	−0.290	–	–	–	–
Jiang et al. (2010)	1956–2004	China	−0.124	−0.149	−0.085	−0.111	−0.151
Fu et al. (2011)	1961–2007	China	−0.130	–	–	–	–
Guo et al. (2011)	1969–2005*	China	−0.180	−0.210	−0.150	−0.160	−0.190
Chen et al. (2013)	1971–2007*	China	−0.170	−0.210	−0.150	−0.160	−0.180

The italicized number indicate the results based on homogenized wind speed data; The number identified by asterisks indicate the results based on data without significant breakpoints but no homogenization

However, compared to the previous studies of regional wind speed trends in China in Table 4, the results here are larger, but with the same order of magnitude as the results of Li et al. (2011). This difference is primarily due to the highest amplitudes having been found in North China; Xu et al. (2006) pointed out that the decreasing trend in annual wind speed series in North China for 1969–2000 was 0.290 m s⁻¹/10a (Table 4), similar to the results here and in Li et al. (2011). The Tianjin region is located in North China, and has the same regional characteristics while also reflecting the territorial climate change. For the processing of research data, the wind speed data have been homogenized systematically by Li et al. (2011) (italicized marking); in the studies of Xu et al. (2006), Guo et al. (2011), and Chen et al. (2013), the authors deleted stations with significant breakpoints detected by metadata or mathematical statistics (asterisk marking) but did not homogenize data. Original data from the studies of Jiang et al. (2010) and Fu et al. (2011) involved no quality control possibly owing to their research needs. Therefore, the differences in results obtained in this study and previously may have been caused by inhomogeneities of the data (Yan et al. 2016). Thus, the trend changes based on adjusted wind speed data here better represent the integration of Beijing, Tianjin, and Hebei.

5.4 Comparison with trend changes in geostrophic wind speed

*Significant at the 95% level

The annual and seasonal geostrophic wind speeds (geowind) at each station are shown in Table 6; they indicate a decreasing trend for the last 50 years (significant at the 95% level are marked with an asterisk). This is consistent with the long-term trend changes in the corresponding near-surface wind speed series (Fig. 8b), illustrating that the wind speed weakened in Tianjin. This physical change may have been related to a weakened system circulation, as suggested by previous studies on near-surface wind speed in China (Xu et al. 2006; Jiang et al. 2010; Guo et al. 2011). The amplitudes of trend change in annual and seasonal geowind at each station were basically the same, the reduction of which was all less obviously than the corresponding near-surface wind speeds (Fig. 8b). However, they were almost the same order of magnitude as trend changes in upper-air wind speed over China (Zhang et al. 2009), and similar to the Kendall trends in 95th and 99th standardized geowind percentile series of the four seasons in the Northeast Atlantic region (Wang et al. 2009). Based on the large-scale circulation background, it is more reasonable that the decreasing trends in near-surface wind speed series have slowed in Tianjin based on the adjusted data (Fig. 8b). However, long-term trend changes in annual and seasonal mean wind speed based on homogenized data at Tianjin (54527) and Ninghe (54529) stations showed a significant decrease in comparison with the raw data (Fig. 8b). This may have been affected by atmospheric circulation, but also by human activities, such as urbanization (Xu et al. 2006; Guo et al. 2011; Azorin-Molina et al. 2014). Results from Sect. 5.3 indicate that the urbanization effect on weakening annual mean wind speed in Tianjin was about 16%, based on the adjusted data. Historical issues, such as observational environment, were also the main factors that affected the discontinuities in the long time series. The wind speed series after adjustment at station 54527 reflected the surrounding environment, including urbanization. There was no detection of local environmental changes at station 54529 owing to the increasingly severe urbanization process. In summary, the real characteristics of region and local climate change in Tianjin were better reflected by the homogenized wind speed series.

5.5 Comparison with the similar data products

Wind speed datasets from NMIC were optimized using the geostrophic wind speed based on sea level pressure from Wan et al. (2010) as a reference series for Eastern China. For the western region, the reference series was constructed using a single station detection combined with adjacent station optimization. The discontinuities in wind speed series caused by station relocation and instrument change were effectively removed using the method RHtest combined with their station metadata. As a result, the regional long-term climate characteristics in China are more accurately reflected based on these surface wind speed data.

Two statistics of standard error (SE) and mean absolute error (MAE) were used here, and the specific methods can be found in Soukissian and Karathanasi (2016).

The values of annual and seasonal SE between the data adjusted here and NMIC were nearly equal, and also for those of MAE in Fig. 9. However, the error amplitudes for MAE (Fig. 9b) were almost less than those of the SE (Fig. 9a), indicating that the average errors between the two kinds of wind speed data were much smaller. The annual and seasonal values of MAE with an error range between 0.2 and 0.4 m/s and probability density up to more than 1.3, and above 1.7 for the autumn data. However, for the SE, the probability density was more than 1.0, and the error range was primarily between 0.3 and 0.6 m/s. In a comparison of the error amplitudes at each station (Table 7), 100% of observation stations had MAE values between 0 and 1.0 m/s for the two kinds of annual or monthly wind speed data, except 92.3% for February data. Correspondingly, there were fewer observation stations with values of SE in the same range, but all reached above 84.6%. The stations with large errors were primarily 54517 and 54527, of which the average annual and monthly values of SE were 1.1 and 1.0 m/s, respectively, and the MAE values were 0.7 and 0.8 m/s. This may have been principally related to the data pre-processing. For NMIC, a number of stations were exchanged before December 31, 1991. This was different than our method, outlined in Sect. 2, which may have caused different breakpoints between the two datasets in the homogenization process, and thus led to some differences in the data adjustment results.

Open image in new window

Table 7

The proportions of errors from 0 to 1.0 m/s between the stations adjusted here and those from CMA; units: percentage

	Jan.	Feb.	Mar.	Apr.	May	Jun.	Jul.	Aug.	Sep.	Oct.	Nov.	Dec.	Ann
SE	84.6	84.6	84.6	92.3	92.3	84.6	92.3	92.3	92.3	92.3	92.3	84.6	92.3
MAE	100	92.3	100	100	100	100	100	100	100	100	100	100	100

Detailed station metadata plays an important role in judging breakpoints in time series for homogenization of climate data (Aguilar et al. 2003). In this study, we pre-processed the data before investigating the historical evolution of each observation station, avoiding judgment errors caused by inhomogeneities from the original data. Moreover, in the adjustment process, to respect the rationality of firsthand observations, we only corrected breakpoints supported by reliable metadata. In NMIC, many breakpoints were removed in the absence of metadata, so the results in this study are relatively conservative. However, in general, 0–1.0 m/s is just a median range breeze on the wind speed scale, with a much smaller error, according to Specification for Ground Meteorological Observation from CMA. Moreover, station relocation and instrument change were also the main reasons for the discontinuities in China monthly mean surface wind speed series based on data descriptions by NMIC, similar to our data. So it was just mentioned in climate data homogeneity studies in China by Li (2016) that different researchers (or team) homogenized on the same climate time series might have some differences, but this difference range should be in a certain limit. Consequently, the homogenized wind speed data for the Tianjin area were highly consistent with those from NMIC; thus, the results in this study should be reliable.