Space-time characteristics of streamflow in Kenya
The major objective of this study was to determine the homogeneous hydrological divisions through space-time analysis of stream flow and to develop the best forecasting and flood prediction regional model. The homogeneous divisions formed the base for further analysis including time-series analysis and model development. The data used in the study included daily stream flow records from 57 sites in Kenya. Daily rainfall and stream flow records for representative stations in the homogeneous regions were used for time series analysis. Daily areal rainfall, daily potential evaporation for Kisumu and Dagoretti stations, and stream flow records for station 3DA02 and 1GD04 for a .period of 16 years were used for two test catchments of Upper Athi and Nyando respectively to calibrate the real-lime forecasting models. The records were subjected to quality control tests using mass and double mass analyses. A non-parametric -runs method was also applied. The quality- controlled data formed a fundamental base of all the analyses in this study. Three methods used in the regionalization included Principal Component Analysis (PCA), the Stu-index and the heterogeneity measure statistics. The visual assessment of L- moment ratio diagrams was also used to further assess the results which were obtained from the Stu-index and the heterogeneity measure methods. Both T and S mode principal component analysis (PCA) solutions were used in this study. The S- mode solutions were used to cluster locations with similar temporal patterns of stream flow based on historical data. T-mode solutions were used to cluster together years with similar spatial stream flow anomalies. The Stu-index, however, is an outlier detection method based on the differences between the arithmetic averages of at site data including or excluding the largest value of record. Regionalization results from the various methods were compared in order to come up with the most realistic hydrological zones for the region of study. Communality concept was further used to select representative locations for each of the homogeneous zones. The patterns derived from the Stu-index method were compared with those derived from the dominant S-mode PCA results. Spectral and cross - spectral analyses were used to investigate the existence of any cyclic variation and to examine the nature of inter - annual recurrences in the seasonal characteristics in the hydrological time series using communality concepts. The core of this study however involved the determination of the best model, which can be used for stream flow forecasting in Kenya. Three types of models were used in this study, namely: Flood frequency distribution models, stream flow - rainfall - area (S-R-A) regression models and deterministic real - time forecasting models. The specific flood frequency models used included the Extreme Value type (EV1), General Extreme Value (GEV) and the Wake by 4 and 5 parameter distributions. The space-time relationships between stream flow, rainfall and catchment area were used to develop the' best S-R-A regression models. The deterministic forecasting models included the simple linear model (SLM), the linear perturbation model (LPM), the variable gain factor linear model (VGFLM) and the soil moisture accounting and routing (SMAR) model. The peA method delineated the region of study into three distinct groups. The first group where the first peA mode was dominant had 40 stations and was located to the western part of the country. The second region which was dominated by the second peA mode was generally located in the central districts had 13 stations, while the third region which was dominated by the third peA mode was concentrated in the eastern towards the coastal parts of Kenya. Only three stations were clustered in this region by the peA method. The Stu-index regionalization method also delineated the stream flow records into three unique categories. The first group consisted of 37 stations were consistent with the peA classification. The other two groups, were however unstable and difficult to separate uniquely. It was, however, noted from the study that although the Stu-index method is based on the outlier detection concepts, its results compared well with the peA solutions. The results from the heterogeneity regi6nalization method based on statistics, H, with those derived from the dispersion in the L-moment ratio diagrams further confirmed the stability of the clustering of the 37 sites above into homogeneous hydrological zone. Unique delineation of the other locations was however relatively more difficult from the regionalization methods which were adopted in this study. The delineated zones for these locations were not stable and varied significantly from one method to another. Regionalization study was therefore restricted largely to this stable homogenous region with a network of 37 stations. Point data were largely used in the other regions. The regional and point data were first subjected to spectral and cross-spectral analyses. No sharp spectral peaks were common in region 1 and 3 for both rainfall and stream flow records. The spectral patterns from most of the time series in these two regions display broad spectral bands. Sharp spectral peaks were however common in region 2. The peaks for rainfall and stream flow compared quite well. The results from cross-spectral analysis however indicated time lag relationship between rainfall and stream flow time series of about zero to two days. Results from the Kolmogorov-Smirnov goodness of fit test, indicated that for the three flood frequency models used, the GEV and the 4 parameter Wakeby were the best distributions. The two models were therefore included in the computation of floods of 50- year and 1OO-year return periods. The results from the application of these models in flood estimates indicated good agreement between the observed and estimated values. Results from stream flow rainfall- area models indicated that in region 1, both catchment area and mean annual rainfall were significant predictors of mean annual flood. The use of catchment area alone underestimated the flood peak, while mean annual rainfall alone overestimated the flood peak. In region 2, however, only the catchment area was found to be a significant predictor of mean annual flood, while use of both mean annual rainfall and catchment area overestimated the flood peak. The use of mean annual rainfall alone also overestimated the flood peak. The results from the deterministic real- time forecasting models based on the model efficiency [R2] indicated that the "COMBINATION" of the simple linear model (SLM), the variable gain factor model (VGFM), and the soil moisture accounting and routing (SMAR) model gave better performance than each of the models individually. The VGFM also gave good results, which were comparable to the results of the 'COMBINATION' model. It was further observed that the autoregressive updating using autoregressive (AR) model gave an improved R2 for both calibration and verification periods. It was also noted from the study that the soil moisture accounting and routing (SMAR) model performed relatively better for a dry catchment of Athi River.
CitationDoctor of Philosophy in Meteorology,
SponsorhipUniversity of Nairobi
Department of Meteorology Faculty of Science University of Nairobi