import numpy as np import pandas as pd # --- replicate the notebook's data prep exactly --- aqi_breakpoints = { 'pm25': [(0,50,0,50),(51,100,51,100),(101,200,101,200),(201,300,201,300)], 'pm10': [(0,50,0,50),(51,100,51,100),(101,250,101,200),(251,350,201,300)], 'co': [(0,1.0,0,50),(1.1,2.0,51,100),(2.1,10.0,101,200),(10.1,17.0,201,300)], } def sub_aqi(c, bps): for il,ih,cl,ch in bps: if cl <= c <= ch: return il if ch==cl else ((ih-il)/(ch-cl))*(c-cl)+il if c < bps[0][2]: return bps[0][0] if c > bps[-1][3]: return bps[-1][1] return np.nan def overall_aqi(row): s=[] for p in ['pm25','pm10','co']: if p in row and not pd.isna(row[p]): s.append(sub_aqi(row[p], aqi_breakpoints[p])) return np.nanmax(s) if s else np.nan df = pd.read_csv('final.csv') df['datetime'] = pd.to_datetime(df['datetime'], format="%d-%m-%Y %H:%M", errors='coerce') df = df.dropna(subset=['datetime']).set_index('datetime').sort_index() df = df[['temp','pm25','pm10','co']].copy() df['calculated_aqi'] = df.apply(overall_aqi, axis=1) df = df[['calculated_aqi','temp','pm25','pm10','co']].copy() df = df.ffill().bfill().dropna() n_ahead = 1 scaler_df = df.shift(n_ahead).rolling(24*14).median() tmp_df = (df / scaler_df).iloc[24*14 + n_ahead:] # this is the RATIO the model was trained on print("=== absolute df (last 24 rows) ===") print(df.tail(24).to_string()) print("\n=== rolling-median denominator (last 24 rows) ===") print(scaler_df.tail(24).to_string()) print("\n=== RATIO tmp_df (last 24 rows) — what the model actually saw ===") print(tmp_df.tail(24).to_string()) print("\n=== ratio min/max per feature across all training data ===") print("min:", tmp_df.min().values) print("max:", tmp_df.max().values) import json, numpy as np from tensorflow.keras.models import load_model from tensorflow.keras.utils import custom_object_scope from tkan import TKAN # reuse df, tmp_df, scaler_df from your script above class MMS: def __init__(s): s.min_=s.max_=s.scale_=None; s.r=(0,1) def load(s,a): s.min_=np.array(a['min_']); s.max_=np.array(a['max_']) s.scale_=np.where(np.array(a['scale_'])==0,1,np.array(a['scale_'])); s.r=tuple(a['minmax_range']) def tf(s,X): return (X-s.min_)/s.scale_*(s.r[1]-s.r[0])+s.r[0] def inv(s,X): return ((X-s.r[0])/(s.r[1]-s.r[0]))*s.scale_+s.min_ isc=MMS(); isc.load(json.load(open('input_scaler_attributes.json'))) tsc=MMS(); tsc.load(json.load(open('target_scaler_attributes.json'))) with custom_object_scope({"TKAN":TKAN,"tkan>TKAN":TKAN}): m=load_model('best_model_TKAN_nahead_1.keras',compile=False) # last 24 ratio rows as model input; the target is the ratio for the NEXT step window = tmp_df.iloc[-25:-1][['calculated_aqi','temp','pm25','pm10','co']].values.reshape(1,24,5) Xs = isc.tf(window) pred_scaled = m.predict(Xs, verbose=0) pred_ratio = tsc.inv(pred_scaled.reshape(1,1,1)).flatten()[0] # denominator for the predicted step = scaler_df AQI at that timestamp denom = scaler_df['calculated_aqi'].iloc[-1] print("predicted ratio :", pred_ratio) print("denominator (median) :", denom) print("recovered AQI :", pred_ratio * denom) print("actual AQI last row :", df['calculated_aqi'].iloc[-1])