{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.12.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceType":"competition","sourceId":35297,"databundleVersionId":3589144}],"dockerImageVersionId":31287,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2026-03-04T06:11:57.992690Z","iopub.execute_input":"2026-03-04T06:11:57.993389Z","iopub.status.idle":"2026-03-04T06:11:58.266008Z","shell.execute_reply.started":"2026-03-04T06:11:57.993360Z","shell.execute_reply":"2026-03-04T06:11:58.265292Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =========================================================\n# 0. IMPORTS\n# =========================================================\nimport os\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\nimport torch\nfrom torch import nn, optim\nfrom torch.utils.data import DataLoader, TensorDataset\n\nfrom sklearn.preprocessing import LabelEncoder\nfrom sklearn.model_selection import KFold\nfrom sklearn.metrics import (\n    roc_auc_score,\n    average_precision_score,\n    precision_score,\n    recall_score,\n    roc_curve,\n    precision_recall_curve\n)\n\nfrom scipy.signal import savgol_filter\n\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nprint(\"Device:\", device)\n\n# =========================================================\n# 1. PARAMETERS\n# =========================================================\nSEQ_LEN = 12\nDIFF_STEPS = 50\nDIFF_T = 15\nBATCH = 512\n\n# =========================================================\n# 2. DATA PREPARATION (UNCHANGED PHYSICS)\n# =========================================================\ndef prepare_data(path):\n\n    df = pd.read_csv(path)\n\n    df[\"timestamp\"] = pd.to_datetime(df[\"timestamp\"], format=\"mixed\", errors=\"coerce\")\n    df = df.dropna(subset=[\"timestamp\"])\n    df = df.sort_values([\"building_id\",\"timestamp\"])\n\n    df[\"meter_log\"] = np.log1p(df[\"meter_reading\"])\n    df[\"hour\"] = df[\"timestamp\"].dt.hour\n    df[\"dow\"] = df[\"timestamp\"].dt.dayofweek\n\n    norm = (\n        df[df[\"anomaly\"]==0]\n        .groupby([\"building_id\",\"dow\",\"hour\"])[\"meter_log\"]\n        .agg([\"median\",\"std\"])\n        .reset_index()\n    )\n\n    norm.columns = [\"building_id\",\"dow\",\"hour\",\"h_med\",\"h_std\"]\n\n    df = df.merge(norm,on=[\"building_id\",\"dow\",\"hour\"],how=\"left\")\n\n    df[\"h_med\"] = df[\"h_med\"].fillna(\n        df.groupby(\"building_id\")[\"meter_log\"].transform(\"median\")\n    )\n\n    df[\"temp_inertia\"] = (\n        df.groupby(\"building_id\")[\"air_temperature\"]\n        .transform(lambda x: x.rolling(6,min_periods=1).mean())\n    )\n\n    df[\"thermal_p\"] = df[\"air_temperature\"] - df[\"temp_inertia\"]\n\n    df[\"shock\"] = df[\"meter_log\"] - df[\"h_med\"]\n    df[\"shock_v\"] = df.groupby(\"building_id\")[\"shock\"].diff().fillna(0)\n    df[\"kinetic_s\"] = df.groupby(\"building_id\")[\"shock_v\"].diff().fillna(0)\n\n    df[\"shock_scaled\"] = df.groupby(\"building_id\")[\"shock\"].transform(\n        lambda x: (x-x.median())/(x.quantile(0.75)-x.quantile(0.25)+1e-6)\n    )\n\n    feats = [\"shock_scaled\",\"thermal_p\",\"kinetic_s\",\"air_temperature\"]\n\n    df[feats] = df[feats].replace([np.inf,-np.inf],0).fillna(0)\n\n    return df,feats\n\n# =========================================================\n# 3. SEQUENCE WINDOWING\n# =========================================================\ndef create_sequences(df,feats):\n\n    X,y,bid = [],[],[]\n\n    for b,g in df.groupby(\"building_id\"):\n\n        g = g.sort_values(\"timestamp\")\n\n        f = g[feats].values\n        lab = g[\"anomaly\"].values\n\n        for i in range(len(g)-SEQ_LEN):\n\n            X.append(f[i:i+SEQ_LEN])\n            y.append(lab[i+SEQ_LEN-1])\n            bid.append(b)\n\n    return (\n        torch.tensor(np.array(X),dtype=torch.float32),\n        np.array(y),\n        np.array(bid)\n    )\n\n# =========================================================\n# 4. LSTM AUTOENCODER\n# =========================================================\nclass LSTMAE(nn.Module):\n\n    def __init__(self,input_dim):\n\n        super().__init__()\n\n        self.encoder = nn.LSTM(\n            input_dim,\n            64,\n            batch_first=True\n        )\n\n        self.decoder = nn.LSTM(\n            64,\n            64,\n            batch_first=True\n        )\n\n        self.output = nn.Linear(64,input_dim)\n\n    def forward(self,x):\n\n        h,_ = self.encoder(x)\n\n        h = h[:,-1,:].unsqueeze(1).repeat(1,x.size(1),1)\n\n        d,_ = self.decoder(h)\n\n        out = self.output(d)\n\n        return torch.nan_to_num(out)\n\n# =========================================================\n# 5. TRAIN + DIFFUSION INFERENCE\n# =========================================================\ndef train_kfold(X,y,bid):\n\n    kf = KFold(n_splits=3,shuffle=True,random_state=42)\n\n    ensemble_scores = np.zeros(len(y))\n\n    alphas = torch.linspace(1e-4,0.02,DIFF_STEPS).to(device)\n\n    for fold,(train_idx,val_idx) in enumerate(kf.split(X)):\n\n        print(f\"\\n🔥 Fold {fold+1}\")\n\n        model = LSTMAE(X.shape[2]).to(device)\n\n        opt = optim.AdamW(model.parameters(),lr=1e-3)\n\n        normal_mask = y[train_idx]==0\n\n        X_train = X[train_idx][normal_mask]\n\n        loader = DataLoader(\n            TensorDataset(X_train),\n            batch_size=BATCH,\n            shuffle=True\n        )\n\n        # TRAIN ONLY ON NORMAL\n        for epoch in range(20):\n\n            model.train()\n\n            for (xb,) in loader:\n\n                xb = xb.to(device)\n\n                recon = model(xb)\n\n                loss = nn.MSELoss()(recon,xb)\n\n                if torch.isfinite(loss):\n\n                    opt.zero_grad()\n                    loss.backward()\n                    torch.nn.utils.clip_grad_norm_(model.parameters(),1.0)\n                    opt.step()\n\n        # ==========================================\n        # INFERENCE WITH DIFFUSION\n        # ==========================================\n\n        model.eval()\n\n        scores=[]\n\n        with torch.no_grad():\n\n            for i in range(0,len(X),BATCH):\n\n                xb = X[i:i+BATCH].to(device)\n\n                target = xb[:,-1,:]\n\n                t = torch.full((xb.size(0),),DIFF_T,device=device)\n\n                noise = torch.randn_like(target)\n\n                alpha = alphas[t].unsqueeze(1)\n\n                noisy_last = torch.sqrt(1-alpha)*target + torch.sqrt(alpha)*noise\n\n                xb_noisy = xb.clone()\n\n                xb_noisy[:,-1,:] = noisy_last\n\n                recon = model(xb_noisy)\n\n                pred = recon[:,-1,:]\n\n                err = (pred-target)**2\n\n                err = (\n                    err[:,0] +\n                    0.75*err[:,2] +\n                    0.5*err[:,1]\n                )\n\n                scores.append(err.cpu())\n\n        fold_scores = torch.cat(scores).numpy()\n\n        fold_scores = np.nan_to_num(fold_scores)\n\n        # Savgol smoothing per building\n        df_tmp = pd.DataFrame({\n            \"score\":fold_scores,\n            \"b\":bid\n        })\n\n        smooth = df_tmp.groupby(\"b\")[\"score\"].transform(\n            lambda x: savgol_filter(x,11,2) if len(x)>11 else x\n        )\n\n        ensemble_scores += smooth/3\n\n    return ensemble_scores\n\n# =========================================================\n# 6. EVALUATION\n# =========================================================\ndef evaluate(scores,y,bid):\n\n    scores = np.nan_to_num(scores)\n\n    auc = roc_auc_score(y,scores)\n    pr = average_precision_score(y,scores)\n\n    thresh = np.percentile(scores,95)\n\n    preds = (scores>thresh).astype(int)\n\n    precision = precision_score(y,preds)\n    recall = recall_score(y,preds)\n\n    fpr = np.sum((preds==1)&(y==0))/np.sum(y==0)\n\n    print(\"\\n🏆 FINAL RESULTS\")\n    print(\"ROC-AUC :\",auc)\n    print(\"PR-AUC  :\",pr)\n    print(\"Precision:\",precision)\n    print(\"Recall:\",recall)\n    print(\"TPR:\",recall)\n    print(\"FPR:\",fpr)\n\n    # ROC\n    fpr_c,tpr_c,_ = roc_curve(y,scores)\n\n    plt.plot(fpr_c,tpr_c)\n    plt.title(\"ROC Curve\")\n    plt.show()\n\n    # PR\n    p_c,r_c,_ = precision_recall_curve(y,scores)\n\n    plt.plot(r_c,p_c)\n    plt.title(\"PR Curve\")\n    plt.show()\n\n    # building example\n    sample = bid[0]\n\n    mask = (bid==sample)\n\n    plt.figure(figsize=(12,5))\n\n    plt.plot(scores[mask],label=\"Predicted\")\n\n    plt.scatter(\n        np.where(y[mask]==1),\n        scores[mask][y[mask]==1],\n        color=\"red\",\n        label=\"Actual Anomaly\"\n    )\n\n    plt.legend()\n    plt.title(f\"Building {sample}\")\n    plt.show()\n\n# =========================================================\n# RUN\n# =========================================================\npath=\"/kaggle/input/competitions/energy-anomaly-detection/train_features.csv\"\n\ndf,feats = prepare_data(path)\n\nX,y,bid = create_sequences(df,feats)\n\nscores = train_kfold(X,y,bid)\n\nevaluate(scores,y,bid)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-03-04T03:59:38.646440Z","iopub.execute_input":"2026-03-04T03:59:38.646876Z","iopub.status.idle":"2026-03-04T04:13:20.220933Z","shell.execute_reply.started":"2026-03-04T03:59:38.646848Z","shell.execute_reply":"2026-03-04T04:13:20.220176Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =========================================================\n# 0. IMPORTS\n# =========================================================\nimport os\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\nimport torch\nfrom torch import nn, optim\nfrom torch.utils.data import DataLoader, TensorDataset\n\nfrom sklearn.model_selection import KFold\nfrom sklearn.metrics import (\n    roc_auc_score,\n    average_precision_score,\n    precision_score,\n    recall_score,\n    roc_curve,\n    precision_recall_curve\n)\n\nfrom scipy.signal import savgol_filter\n\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nprint(\"Device:\", device)\n\n# =========================================================\n# PARAMETERS\n# =========================================================\nSEQ_LEN = 12\nDIFF_STEPS = 50\nDIFF_T = 15\nBATCH = 512\n\n# =========================================================\n# DATA PREPARATION (PHYSICS FEATURES)\n# =========================================================\ndef prepare_data(path):\n\n    df = pd.read_csv(path)\n\n    df[\"timestamp\"] = pd.to_datetime(df[\"timestamp\"], format=\"mixed\", errors=\"coerce\")\n    df = df.dropna(subset=[\"timestamp\"])\n    df = df.sort_values([\"building_id\",\"timestamp\"])\n\n    df[\"meter_log\"] = np.log1p(df[\"meter_reading\"])\n    df[\"hour\"] = df[\"timestamp\"].dt.hour\n    df[\"dow\"] = df[\"timestamp\"].dt.dayofweek\n\n    norm = (\n        df[df[\"anomaly\"]==0]\n        .groupby([\"building_id\",\"dow\",\"hour\"])[\"meter_log\"]\n        .agg([\"median\",\"std\"])\n        .reset_index()\n    )\n\n    norm.columns = [\"building_id\",\"dow\",\"hour\",\"h_med\",\"h_std\"]\n\n    df = df.merge(norm,on=[\"building_id\",\"dow\",\"hour\"],how=\"left\")\n\n    df[\"h_med\"] = df[\"h_med\"].fillna(\n        df.groupby(\"building_id\")[\"meter_log\"].transform(\"median\")\n    )\n\n    df[\"temp_inertia\"] = (\n        df.groupby(\"building_id\")[\"air_temperature\"]\n        .transform(lambda x: x.rolling(6,min_periods=1).mean())\n    )\n\n    df[\"thermal_p\"] = df[\"air_temperature\"] - df[\"temp_inertia\"]\n\n    df[\"shock\"] = df[\"meter_log\"] - df[\"h_med\"]\n    df[\"shock_v\"] = df.groupby(\"building_id\")[\"shock\"].diff().fillna(0)\n    df[\"kinetic_s\"] = df.groupby(\"building_id\")[\"shock_v\"].diff().fillna(0)\n\n    df[\"shock_scaled\"] = df.groupby(\"building_id\")[\"shock\"].transform(\n        lambda x:(x-x.median())/(x.quantile(0.75)-x.quantile(0.25)+1e-6)\n    )\n\n    feats = [\"shock_scaled\",\"thermal_p\",\"kinetic_s\",\"air_temperature\"]\n\n    df[feats] = df[feats].replace([np.inf,-np.inf],0).fillna(0)\n\n    return df,feats\n\n# =========================================================\n# SEQUENCE WINDOWING\n# =========================================================\ndef create_sequences(df,feats):\n\n    X,y,bid = [],[],[]\n\n    for b,g in df.groupby(\"building_id\"):\n\n        g = g.sort_values(\"timestamp\")\n\n        f = g[feats].values\n        lab = g[\"anomaly\"].values\n\n        for i in range(len(g)-SEQ_LEN):\n\n            X.append(f[i:i+SEQ_LEN])\n            y.append(lab[i+SEQ_LEN-1])\n            bid.append(b)\n\n    return (\n        torch.tensor(np.array(X),dtype=torch.float32),\n        np.array(y),\n        np.array(bid)\n    )\n\n# =========================================================\n# LSTM AUTOENCODER\n# =========================================================\nclass LSTMAE(nn.Module):\n\n    def __init__(self,input_dim):\n\n        super().__init__()\n\n        self.encoder = nn.LSTM(input_dim,64,batch_first=True)\n        self.decoder = nn.LSTM(64,64,batch_first=True)\n\n        self.output = nn.Linear(64,input_dim)\n\n    def forward(self,x):\n\n        h,_ = self.encoder(x)\n\n        h = h[:,-1,:].unsqueeze(1).repeat(1,x.size(1),1)\n\n        d,_ = self.decoder(h)\n\n        out = self.output(d)\n\n        return torch.nan_to_num(out)\n\n# =========================================================\n# TRAIN + DIFFUSION INFERENCE\n# =========================================================\ndef train_kfold(X,y,bid):\n\n    kf = KFold(n_splits=3,shuffle=True,random_state=42)\n\n    ensemble_scores = np.zeros(len(y))\n\n    alphas = torch.linspace(1e-4,0.02,DIFF_STEPS).to(device)\n\n    for fold,(train_idx,val_idx) in enumerate(kf.split(X)):\n\n        print(f\"\\n🔥 Fold {fold+1}\")\n\n        model = LSTMAE(X.shape[2]).to(device)\n\n        opt = optim.AdamW(model.parameters(),lr=1e-3)\n\n        normal_mask = y[train_idx]==0\n\n        X_train = X[train_idx][normal_mask]\n\n        loader = DataLoader(\n            TensorDataset(X_train),\n            batch_size=BATCH,\n            shuffle=True\n        )\n\n        # ==========================\n        # TRAIN ONLY ON NORMAL DATA\n        # ==========================\n        for epoch in range(20):\n\n            model.train()\n\n            for (xb,) in loader:\n\n                xb = xb.to(device)\n\n                recon = model(xb)\n\n                loss = nn.MSELoss()(recon,xb)\n\n                if torch.isfinite(loss):\n\n                    opt.zero_grad()\n                    loss.backward()\n                    torch.nn.utils.clip_grad_norm_(model.parameters(),1.0)\n                    opt.step()\n\n        # ==========================\n        # INFERENCE\n        # ==========================\n        model.eval()\n\n        scores=[]\n\n        with torch.no_grad():\n\n            for i in range(0,len(X),BATCH):\n\n                xb = X[i:i+BATCH].to(device)\n\n                target = xb[:,-1,:]\n\n                t = torch.full((xb.size(0),),DIFF_T,device=device)\n\n                noise = torch.randn_like(target)\n\n                alpha = alphas[t].unsqueeze(1)\n\n                noisy_last = torch.sqrt(1-alpha)*target + torch.sqrt(alpha)*noise\n\n                xb_noisy = xb.clone()\n                xb_noisy[:,-1,:] = noisy_last\n\n                recon = model(xb_noisy)\n\n                pred = recon[:,-1,:]\n\n                err = (pred-target)**2\n\n                err = (\n                    1.2*err[:,0] +    # shock_scaled\n                    0.8*err[:,2] +    # kinetic_s\n                    0.4*err[:,1]      # thermal_p\n                )\n\n                scores.append(err.cpu())\n\n        fold_scores = torch.cat(scores).numpy()\n        fold_scores = np.nan_to_num(fold_scores)\n\n        # =====================================================\n        # PER-BUILDING NORMALIZATION (FIX)\n        # =====================================================\n        df_tmp = pd.DataFrame({\n            \"score\":fold_scores,\n            \"b\":bid\n        })\n\n        df_tmp[\"score\"] = df_tmp.groupby(\"b\")[\"score\"].transform(\n            lambda x:(x-x.mean())/(x.std()+1e-6)\n        )\n\n        # =====================================================\n        # SMOOTHING\n        # =====================================================\n        smooth = df_tmp.groupby(\"b\")[\"score\"].transform(\n            lambda x: savgol_filter(x,11,2) if len(x)>11 else x\n        )\n\n        ensemble_scores += smooth/3\n\n    return ensemble_scores\n\n# =========================================================\n# EVALUATION\n# =========================================================\ndef evaluate(scores,y,bid):\n\n    scores = np.nan_to_num(scores)\n\n    auc = roc_auc_score(y,scores)\n    pr = average_precision_score(y,scores)\n\n    thresh = np.percentile(scores,95)\n\n    preds = (scores>thresh).astype(int)\n\n    precision = precision_score(y,preds)\n    recall = recall_score(y,preds)\n\n    fpr = np.sum((preds==1)&(y==0))/np.sum(y==0)\n\n    print(\"\\n🏆 FINAL RESULTS\")\n    print(\"ROC-AUC :\",auc)\n    print(\"PR-AUC  :\",pr)\n    print(\"Precision:\",precision)\n    print(\"Recall:\",recall)\n    print(\"TPR:\",recall)\n    print(\"FPR:\",fpr)\n\n    # ROC Curve\n    fpr_c,tpr_c,_ = roc_curve(y,scores)\n    plt.plot(fpr_c,tpr_c)\n    plt.title(\"ROC Curve\")\n    plt.show()\n\n    # PR Curve\n    p_c,r_c,_ = precision_recall_curve(y,scores)\n    plt.plot(r_c,p_c)\n    plt.title(\"PR Curve\")\n    plt.show()\n\n    # Building visualization\n    sample = bid[0]\n\n    mask = bid==sample\n\n    plt.figure(figsize=(12,5))\n    plt.plot(scores[mask],label=\"Predicted\")\n\n    plt.scatter(\n        np.where(y[mask]==1),\n        scores[mask][y[mask]==1],\n        color=\"red\",\n        label=\"Actual Anomaly\"\n    )\n\n    plt.legend()\n    plt.title(f\"Building {sample}\")\n    plt.show()\n\n# =========================================================\n# RUN\n# =========================================================\npath=\"/kaggle/input/competitions/energy-anomaly-detection/train_features.csv\"\n\ndf,feats = prepare_data(path)\n\nX,y,bid = create_sequences(df,feats)\n\nscores = train_kfold(X,y,bid)\n\nevaluate(scores,y,bid)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-03-04T06:12:04.575222Z","iopub.execute_input":"2026-03-04T06:12:04.575955Z","iopub.status.idle":"2026-03-04T06:26:51.781365Z","shell.execute_reply.started":"2026-03-04T06:12:04.575928Z","shell.execute_reply":"2026-03-04T06:26:51.780697Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =========================================================\n# IMPORTS\n# =========================================================\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\nimport torch\nfrom torch import nn, optim\nfrom torch.utils.data import DataLoader, TensorDataset\n\nfrom sklearn.model_selection import KFold\nfrom sklearn.metrics import (\n    roc_auc_score,\n    average_precision_score,\n    precision_score,\n    recall_score,\n    roc_curve,\n    precision_recall_curve\n)\n\nfrom scipy.signal import savgol_filter\n\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nprint(\"Device:\", device)\n\n# =========================================================\n# PARAMETERS\n# =========================================================\nSEQ_LEN = 12\nDIFF_STEPS = 50\nDIFF_T = 15\nBATCH = 512\n\n# =========================================================\n# DATA PREPARATION (PHYSICS FEATURES)\n# =========================================================\ndef prepare_data(path):\n\n    df = pd.read_csv(path)\n\n    df[\"timestamp\"] = pd.to_datetime(df[\"timestamp\"], format=\"mixed\", errors=\"coerce\")\n    df = df.dropna(subset=[\"timestamp\"])\n    df = df.sort_values([\"building_id\",\"timestamp\"])\n\n    df[\"meter_log\"] = np.log1p(df[\"meter_reading\"])\n    df[\"hour\"] = df[\"timestamp\"].dt.hour\n    df[\"dow\"] = df[\"timestamp\"].dt.dayofweek\n\n    norm = (\n        df[df[\"anomaly\"]==0]\n        .groupby([\"building_id\",\"dow\",\"hour\"])[\"meter_log\"]\n        .agg([\"median\",\"std\"])\n        .reset_index()\n    )\n\n    norm.columns = [\"building_id\",\"dow\",\"hour\",\"h_med\",\"h_std\"]\n\n    df = df.merge(norm,on=[\"building_id\",\"dow\",\"hour\"],how=\"left\")\n\n    df[\"h_med\"] = df[\"h_med\"].fillna(\n        df.groupby(\"building_id\")[\"meter_log\"].transform(\"median\")\n    )\n\n    df[\"temp_inertia\"] = (\n        df.groupby(\"building_id\")[\"air_temperature\"]\n        .transform(lambda x: x.rolling(6,min_periods=1).mean())\n    )\n\n    df[\"thermal_p\"] = df[\"air_temperature\"] - df[\"temp_inertia\"]\n\n    df[\"shock\"] = df[\"meter_log\"] - df[\"h_med\"]\n    df[\"shock_v\"] = df.groupby(\"building_id\")[\"shock\"].diff().fillna(0)\n    df[\"kinetic_s\"] = df.groupby(\"building_id\")[\"shock_v\"].diff().fillna(0)\n\n    df[\"shock_scaled\"] = df.groupby(\"building_id\")[\"shock\"].transform(\n        lambda x:(x-x.median())/(x.quantile(0.75)-x.quantile(0.25)+1e-6)\n    )\n\n    feats = [\"shock_scaled\",\"thermal_p\",\"kinetic_s\",\"air_temperature\"]\n\n    df[feats] = df[feats].replace([np.inf,-np.inf],0).fillna(0)\n\n    return df,feats\n\n# =========================================================\n# SEQUENCE WINDOWING\n# =========================================================\ndef create_sequences(df,feats):\n\n    X,y,bid = [],[],[]\n\n    for b,g in df.groupby(\"building_id\"):\n\n        g = g.sort_values(\"timestamp\")\n\n        f = g[feats].values\n        lab = g[\"anomaly\"].values\n\n        for i in range(len(g)-SEQ_LEN):\n\n            X.append(f[i:i+SEQ_LEN])\n            y.append(lab[i+SEQ_LEN-1])\n            bid.append(b)\n\n    return (\n        torch.tensor(np.array(X),dtype=torch.float32),\n        np.array(y),\n        np.array(bid)\n    )\n\n# =========================================================\n# LSTM AUTOENCODER\n# =========================================================\nclass LSTMAE(nn.Module):\n\n    def __init__(self,input_dim):\n\n        super().__init__()\n\n        self.encoder = nn.LSTM(input_dim,64,batch_first=True)\n        self.decoder = nn.LSTM(64,64,batch_first=True)\n\n        self.output = nn.Linear(64,input_dim)\n\n    def forward(self,x):\n\n        h,_ = self.encoder(x)\n\n        h = h[:,-1,:].unsqueeze(1).repeat(1,x.size(1),1)\n\n        d,_ = self.decoder(h)\n\n        out = self.output(d)\n\n        return torch.nan_to_num(out)\n\n# =========================================================\n# TRAIN + DIFFUSION INFERENCE\n# =========================================================\ndef train_kfold(X,y,bid):\n\n    kf = KFold(n_splits=3,shuffle=True,random_state=42)\n\n    ensemble_scores = np.zeros(len(y))\n\n    alphas = torch.linspace(1e-4,0.02,DIFF_STEPS).to(device)\n\n    for fold,(train_idx,val_idx) in enumerate(kf.split(X)):\n\n        print(f\"\\n🔥 Fold {fold+1}\")\n\n        model = LSTMAE(X.shape[2]).to(device)\n\n        opt = optim.AdamW(model.parameters(),lr=1e-3)\n\n        normal_mask = y[train_idx]==0\n        X_train = X[train_idx][normal_mask]\n\n        loader = DataLoader(\n            TensorDataset(X_train),\n            batch_size=BATCH,\n            shuffle=True\n        )\n\n        # TRAIN\n        for epoch in range(20):\n\n            model.train()\n\n            for (xb,) in loader:\n\n                xb = xb.to(device)\n\n                recon = model(xb)\n\n                loss = nn.MSELoss()(recon,xb)\n\n                if torch.isfinite(loss):\n\n                    opt.zero_grad()\n                    loss.backward()\n                    torch.nn.utils.clip_grad_norm_(model.parameters(),1.0)\n                    opt.step()\n\n        # INFERENCE\n        model.eval()\n        scores=[]\n\n        with torch.no_grad():\n\n            for i in range(0,len(X),BATCH):\n\n                xb = X[i:i+BATCH].to(device)\n\n                target = xb[:,-1,:]\n\n                t = torch.full((xb.size(0),),DIFF_T,device=device)\n\n                noise = torch.randn_like(target)\n\n                alpha = alphas[t].unsqueeze(1)\n\n                noisy_last = torch.sqrt(1-alpha)*target + torch.sqrt(alpha)*noise\n\n                xb_noisy = xb.clone()\n                xb_noisy[:,-1,:] = noisy_last\n\n                recon = model(xb_noisy)\n\n                pred = recon[:,-1,:]\n\n                err = (pred-target)**2\n\n                err = (\n                    1.2*err[:,0] +\n                    0.8*err[:,2] +\n                    0.4*err[:,1]\n                )\n                    \n                scores.append(err.cpu())\n\n        fold_scores = torch.cat(scores).numpy()\n        fold_scores = np.nan_to_num(fold_scores)\n\n        df_tmp = pd.DataFrame({\n            \"score\":fold_scores,\n            \"b\":bid\n        })\n\n        # per-building normalization\n        df_tmp[\"score\"] = df_tmp.groupby(\"b\")[\"score\"].transform(\n            lambda x:(x-x.mean())/(x.std()+1e-6)\n        )\n\n        # smoothing\n        smooth = df_tmp.groupby(\"b\")[\"score\"].transform(\n            lambda x: savgol_filter(x,11,2) if len(x)>11 else x\n        )\n\n        ensemble_scores += smooth/3\n\n    # =====================================================\n    # PR BOOST TRICKS\n    # =====================================================\n    ensemble_scores = np.sign(ensemble_scores)*np.log1p(np.abs(ensemble_scores))\n    ensemble_scores = pd.Series(ensemble_scores).rank(pct=True).values\n\n    return ensemble_scores\n\n# =========================================================\n# EVALUATION\n# =========================================================\ndef evaluate(scores,y,bid):\n\n    auc = roc_auc_score(y,scores)\n    pr = average_precision_score(y,scores)\n\n    thresh = np.percentile(scores,95)\n    preds = (scores>thresh).astype(int)\n\n    precision = precision_score(y,preds)\n    recall = recall_score(y,preds)\n\n    fpr = np.sum((preds==1)&(y==0))/np.sum(y==0)\n\n    print(\"\\n🏆 FINAL RESULTS\")\n    print(\"ROC-AUC :\",auc)\n    print(\"PR-AUC  :\",pr)\n    print(\"Precision:\",precision)\n    print(\"Recall:\",recall)\n    print(\"TPR:\",recall)\n    print(\"FPR:\",fpr)\n\n    fpr_c,tpr_c,_ = roc_curve(y,scores)\n    plt.plot(fpr_c,tpr_c)\n    plt.title(\"ROC Curve\")\n    plt.show()\n\n    p_c,r_c,_ = precision_recall_curve(y,scores)\n    plt.plot(r_c,p_c)\n    plt.title(\"PR Curve\")\n    plt.show()\n\n    sample = bid[0]\n    mask = bid==sample\n\n    plt.figure(figsize=(12,5))\n    plt.plot(scores[mask],label=\"Predicted\")\n    plt.scatter(np.where(y[mask]==1),scores[mask][y[mask]==1],color=\"red\")\n    plt.legend()\n    plt.title(f\"Building {sample}\")\n    plt.show()\n\n# =========================================================\n# RUN\n# =========================================================\npath=\"/kaggle/input/competitions/energy-anomaly-detection/train_features.csv\"\n\ndf,feats = prepare_data(path)\nX,y,bid = create_sequences(df,feats)\n\nscores = train_kfold(X,y,bid)\n\nevaluate(scores,y,bid)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-03-04T08:40:38.260185Z","iopub.execute_input":"2026-03-04T08:40:38.260492Z","iopub.status.idle":"2026-03-04T08:54:34.451669Z","shell.execute_reply.started":"2026-03-04T08:40:38.260467Z","shell.execute_reply":"2026-03-04T08:54:34.450870Z"}},"outputs":[],"execution_count":null}]}