Skip to content

Latest commit

 

History

History
344 lines (317 loc) · 10.6 KB

File metadata and controls

344 lines (317 loc) · 10.6 KB

Final Feature Stack

This file documents the exact final feature stack used by the best FloodGraphFlow-XGB submission path.

  • Base engineered features before pruning: `231
  • Appended Stage A stacked features: 31
  • Final regressor input dimension: 262

Base Features (231)

  1. rain_sum_w1
  2. rain_sum_w2
  3. rain_sum_w4
  4. rain_sum_w8
  5. rain_sum_w16
  6. rain_sum_w32
  7. rain_sum_w64
  8. rain_max_w1
  9. rain_max_w2
  10. rain_max_w4
  11. rain_max_w8
  12. rain_max_w16
  13. rain_max_w32
  14. rain_max_w64
  15. rain_ewm_a0p9
  16. rain_ewm_a0p97
  17. rain_ewm_a0p99
  18. warm_wl_first
  19. warm_wl_last
  20. warm_wl_mean
  21. warm_wl_std
  22. warm_wl_slope
  23. warm_dwl_last
  24. warm_d2wl_last
  25. warm_dwl_abs_mean_short
  26. warm_dwl_abs_mean_long
  27. warm_dwl_burst_ratio
  28. position_x
  29. position_y
  30. depth
  31. invert_elevation
  32. surface_elevation
  33. base_area
  34. rel_elev_local_vs_neighbors
  35. sink_depth_rel_neighbors
  36. cap_fullness_1d
  37. cap_headroom_1d
  38. cap_surface_gap_1d
  39. cap_pond_depth_2d
  40. cap_pond_norm_2d
  41. cap_near_1d_0p8
  42. cap_near_1d_0p9
  43. cap_near_2d_0p8
  44. cap_near_2d_0p9
  45. pipe_has_outflow
  46. pipe_min_outflow_diameter
  47. pipe_mean_outflow_diameter
  48. t_rel
  49. t_rel_norm
  50. lead_bin_0_35
  51. lead_bin_36_71
  52. lead_bin_72_107
  53. lead_bin_108_223
  54. lead_bin_224_plus
  55. rain_burst_ratio_sum
  56. rain_burst_ratio_max
  57. rain_since_last
  58. rain_since_peak
  59. future_rain_sum_1
  60. future_rain_max_1
  61. future_rain_sum_2
  62. future_rain_max_2
  63. future_rain_sum_4
  64. future_rain_max_4
  65. future_rain_sum_8
  66. future_rain_max_8
  67. future_rain_sum_16
  68. future_rain_max_16
  69. future_rain_sum_32
  70. future_rain_max_32
  71. future_rain_sum_64
  72. future_rain_max_64
  73. endpoint_is_inlet_1d
  74. endpoint_is_outfall_1d
  75. endpoint_is_boundary_1d
  76. endpoint_inlet_x_future_rain8
  77. endpoint_inlet_x_future_rain24
  78. endpoint_inlet_x_future_rain32
  79. endpoint_outfall_x_ds_lock
  80. endpoint_outfall_x_interface_drain_pressure
  81. endpoint_boundary_x_drain_gate
  82. endpoint_boundary_x_late_gate
  83. endpoint_boundary_x_tnorm
  84. endpoint_inlet_x_future_rain64
  85. endpoint_inlet_x_rain_ewm_0p99
  86. endpoint_inlet_x_basin_mass_deficit_ratio
  87. endpoint_inlet_x_node_prior_q95
  88. routing_rain_lag_1
  89. routing_rain_lag_2
  90. routing_rain_lag_4
  91. routing_rain_lag_8
  92. routing_rain_lag_12
  93. routing_rain_lag_24
  94. routing_rain_lag_36
  95. routing_rain_lag_48
  96. routing_rain_dr1
  97. routing_rain_slope4
  98. routing_rain_slope8
  99. routing_since_global_peak
  100. routing_since_last_rain
  101. routing_cum_since_calm
  102. routing_area_x_rain_now
  103. routing_area_x_rain_short
  104. routing_flow_accumulation_x_rain_now
  105. routing_flow_accumulation_x_rain_short
  106. routing_centroid_elevation_x_rain_now
  107. routing_centroid_elevation_x_rain_short
  108. upstream_rain_sum_w4
  109. upstream_rain_mean_w4
  110. upstream_rain_hop2_sum_w4
  111. upstream_rain_hop2_mean_w4
  112. upstream_rain_sum_w8
  113. upstream_rain_mean_w8
  114. upstream_rain_hop2_sum_w8
  115. upstream_rain_hop2_mean_w8
  116. upstream_rain_sum_w16
  117. upstream_rain_mean_w16
  118. upstream_rain_hop2_sum_w16
  119. upstream_rain_hop2_mean_w16
  120. upstream_rain_sum_w32
  121. upstream_rain_mean_w32
  122. upstream_rain_hop2_sum_w32
  123. upstream_rain_hop2_mean_w32
  124. interaction_rain_now_x_warm_slope
  125. interaction_rain_now_x_warm_abs_dy_mean_long
  126. interaction_depth_x_rain_now
  127. interaction_invert_elevation_x_rain_now
  128. interaction_surface_elevation_x_rain_now
  129. hydro_up_cap
  130. hydro_dn_cap
  131. hydro_dn_cap_min
  132. hydro_up_flow
  133. hydro_dn_flow
  134. hydro_up_wl_mean
  135. hydro_pressure_mean_1d2d
  136. hydro_fill_pct
  137. hydro_freeboard
  138. hydro_base_area
  139. hydro_area
  140. hydro_flow_accumulation
  141. fillrain_rain_now_x_fill_pct
  142. fillrain_rain_short_x_fill_pct
  143. fillrain_future_rain16_x_fill_pct
  144. fillrain_rain_now_div_freeboard
  145. fillrain_rain_short_div_freeboard
  146. fillrain_future_rain16_div_freeboard
  147. tx_rain_now_fill_log1p
  148. tx_rain_short_fill_log1p
  149. tx_future_rain4_fill_log1p
  150. tx_rain_now_div_free_log1p
  151. tx_rain_short_div_free_log1p
  152. tx_future_rain16_div_free_log1p
  153. tx_dvdh_area_ratio_log1p
  154. tx_store_delta_asinh
  155. tx_qhist_fast_qin_upsum_asinh
  156. tx_qhist_slow_qin_upsum_asinh
  157. tx_qhist_fast_qout_upsum_asinh
  158. tx_targeted_freeboard_warm_asinh
  159. tx_freeboard_asinh
  160. tx_headroom_1d_asinh
  161. tx_dvdh_abs_dh_asinh
  162. tx_node_wl_over_basin_mean_stable_asinh
  163. zae_base_zero
  164. zae_endpoint_zero
  165. zae_base_zero_x_release
  166. zae_base_zero_x_unlock
  167. zae_endpoint_zero_x_rain8
  168. zae_endpoint_zero_x_rain32
  169. zae_endpoint_zero_x_freeboard
  170. targeted_indeg_1d
  171. targeted_outdeg_1d
  172. targeted_up_len_1d
  173. targeted_dn_len_1d
  174. targeted_sink_depth_2d
  175. targeted_is_surcharged_warm
  176. targeted_freeboard_warm
  177. targeted_lag8_rain_x_up_len_1d
  178. targeted_has_sink_depth
  179. dvdh_live_area
  180. dvdh_area_ratio
  181. dvdh_abs_dh
  182. global_outlet_hops_norm
  183. global_elev_above_outfall
  184. basin_mean_wl_lag9
  185. basin_rain24
  186. node_wl_over_basin_mean
  187. twi_proxy
  188. spi_proxy
  189. slope_proxy
  190. log1p_area_proxy
  191. routingpot_outlet_hops_norm
  192. routingpot_upstream_area_frac
  193. routingpot_hierarchy_proxy
  194. routingpot_inout_ratio
  195. routingpot_travel_impedance
  196. routingpot_node_len_mean
  197. hand_surface
  198. hand_centroid
  199. hand_invert
  200. hand_nearest_drain_inv
  201. hand_wl_minus_nearest_drain_inv
  202. basin_mass_deficit
  203. basin_mass_deficit_ratio
  204. basin_storage_now
  205. downstream_full_frac
  206. downstream_headroom_mean
  207. downstream_wl_mean
  208. downstream_lock_pressure
  209. subcatch_mass_deficit
  210. subcatch_mass_deficit_ratio
  211. subcatch_storage_now
  212. mass_node_deficit
  213. mass_up1_deficit_mean
  214. mass_up2_deficit_mean
  215. mass_node_minus_basin_deficit
  216. mass_up1_minus_basin_deficit
  217. mass_up2_minus_basin_deficit
  218. node_prior_max_wl
  219. node_prior_q95_wl
  220. node_prior_flood_freq
  221. drain_prior_dry_drop_q50
  222. drain_prior_dry_drop_q90
  223. drain_prior_dry_drop_rate
  224. drain_prior_dry_recession_slope
  225. uphist_ema_fast_upsum
  226. uphist_ema_slow_upsum
  227. uphist_ema_fast_upmean
  228. uphist_ema_slow_upmean
  229. uphist_ema_fastslow_ratio
  230. uphist_ema_outfall_slow
  231. interaction_base_area_x_rain_now

Stage A Appended Features (31)

  1. qnet_pred
  2. qnet_cumsum
  3. qnet_rollsum_3
  4. qnet_rollsum_8
  5. qnet_rollsum_24
  6. qnet_lag_1
  7. qnet_lag_2
  8. qnet_lag_4
  9. qhat_upstream_sum
  10. qhat_upstream_mean
  11. qhat_up_down_imbalance
  12. qhat_upstream_hop2_sum
  13. qhat_upstream_hop2_mean
  14. qnet_phys_baseline_wl
  15. qin_pred
  16. qout_pred
  17. qin_cumsum
  18. qin_rollsum_3
  19. qin_rollsum_8
  20. qin_rollsum_24
  21. qin_lag_1
  22. qin_lag_2
  23. qin_lag_4
  24. qout_cumsum
  25. qout_rollsum_3
  26. qout_rollsum_8
  27. qout_rollsum_24
  28. qout_lag_1
  29. qout_lag_2
  30. qout_lag_4
  31. aux_peak_within_24

Derivation Map

Warm-start rainfall history

  • Derived from the observed warm-start rainfall sequence only.
  • rain_sum_* are rolling sums.
  • rain_max_* are rolling maxima.
  • rain_ewm_* are exponentially weighted moving summaries.

Warm-start water-level summaries

  • warm_wl_* summarize the observed warm-start water-level history.
  • warm_d* summarize first and second differences over the same warm-start window.

Static and geometry context

  • position_x, position_y, depth, invert_elevation, surface_elevation, and base_area are direct static inputs.
  • rel_elev_* compare a node’s elevation to its undirected neighbors.
  • cap_* encode fullness, headroom, surface gap, and 1D/2D near-capacity indicators.
  • pipe_* are static 1D outflow bottleneck descriptors.
  • global_*, twi_*, spi_*, routingpot_*, and hand_* are static topology/terrain/drainage proxies.

Time and forcing context

  • t_rel, t_rel_norm, and lead_bin_* condition the model on forecast horizon.
  • rain_burst_* capture burstiness and recency.
  • future_rain_* are leak-safe known rainfall forcing windows.
  • endpoint_* specialize forcing and lockup context for inlet/outfall/boundary nodes.
  • routing_* encode rainfall lags, rise-rate, recency, and a few routing-style static interactions.
  • upstream_rain_* aggregate 1-hop and 2-hop upstream rain summaries.

Interaction features

  • interaction_* are simple nonlinear rain-by-state or rain-by-static products.
  • fillrain_* combine rainfall with fill ratio and freeboard.

Hydraulic state and physics-aware features

  • hydro_* encode local/upstream/downstream hydraulic context from warm-start state, conveyance, and geometry.
  • targeted_* are hand-built topology and surcharge-sensitive physics proxies.
  • dvdh_* approximate local storage-response behavior.
  • basin_mean_wl_lag9, basin_rain24, and node_wl_over_basin_mean provide basin-scale context.

Mass-balance features

  • basin_mass_deficit* are basin-scale storage deficit proxies.
  • subcatch_mass_deficit* are the same idea at subcatchment scale.
  • mass_*deficit* compare node, 1-hop upstream, 2-hop upstream, and basin imbalance levels.
  • These are the main mass-balance-style features that drove the strong mid-branch score drop.

Priors and historical regime features

  • node_prior_* are train-split node priors from historical maximum water level, q95 water level, and exceedance frequency.
  • drain_prior_* are train-split dry/recession priors.
  • uphist_ema_* summarize upstream warm-start history with fast and slow EMA channels.

Transform cleanup v3

  • tx_* features are stabilized log1p or asinh transforms of otherwise brittle hydraulic ratios or interactions.
  • zae_* features explicitly handle zero-area endpoint edge cases.
  • This cleanup block is the main feature difference between final public Model_2 and Model_1.

Stage A stacked channels

  • qnet_* are out-of-fold net-flow surrogate predictions plus cumulative, rolling, and lag summaries.
  • qhat_* summarize those predicted net flows over the drainage graph.
  • qin_* and qout_* are out-of-fold inflow and outflow surrogate predictions plus state summaries.
  • aux_peak_within_24 is the appended auxiliary event target prediction.