ems-core/backend/app/services/ai_ops.py

"""AI运维智能体服务 - 设备健康评分、异常检测、诊断智能、预测性维护、运营洞察

Inspired by Envision's "构网智能体" concept: self-sensing, self-adapting, self-evolving
intelligent agents for energy asset management.
"""
import logging
import random
import math
from datetime import datetime, timezone, timedelta
from sqlalchemy import select, func, and_, desc
from sqlalchemy.ext.asyncio import AsyncSession
from app.models.device import Device
from app.models.energy import EnergyData, EnergyDailySummary
from app.models.alarm import AlarmEvent
from app.models.ai_ops import (
    DeviceHealthScore, AnomalyDetection, DiagnosticReport,
    MaintenancePrediction, OpsInsight,
)

logger = logging.getLogger("ai_ops")

# ── Device type configurations ──────────────────────────────────────

DEVICE_RATED_EFFICIENCY = {
    "pv_inverter": {"metric": "power", "rated_cop": None, "temp_range": (20, 60)},
    "heat_pump": {"metric": "cop", "rated_cop": 3.5, "temp_range": (30, 55)},
    "meter": {"metric": "power_factor", "rated_cop": None, "temp_range": None},
    "sensor": {"metric": "temperature", "rated_cop": None, "temp_range": (18, 28)},
    "heat_meter": {"metric": "heat_power", "rated_cop": None, "temp_range": None},
}

HEALTH_WEIGHTS = {
    "power_stability": 0.20,
    "efficiency": 0.25,
    "alarm_frequency": 0.20,
    "uptime": 0.20,
    "temperature": 0.15,
}


# ── Health Score Calculation ────────────────────────────────────────

async def calculate_device_health(
    session: AsyncSession, device: Device, now: datetime | None = None
) -> DeviceHealthScore:
    """Calculate health score (0-100) for a single device based on weighted factors."""
    now = now or datetime.now(timezone.utc)
    factors = {}

    # Factor 1: Power output stability (std_dev of power over last 24h)
    factors["power_stability"] = await _calc_power_stability(session, device.id, now)

    # Factor 2: Efficiency / COP
    factors["efficiency"] = await _calc_efficiency_score(session, device, now)

    # Factor 3: Alarm frequency (last 7 days)
    factors["alarm_frequency"] = await _calc_alarm_frequency_score(session, device.id, now)

    # Factor 4: Uptime (last 30 days)
    factors["uptime"] = await _calc_uptime_score(session, device, now)

    # Factor 5: Temperature
    factors["temperature"] = await _calc_temperature_score(session, device, now)

    # Weighted average
    health_score = sum(
        factors[k] * HEALTH_WEIGHTS[k] for k in HEALTH_WEIGHTS
    )
    health_score = max(0, min(100, round(health_score, 1)))

    # Status
    if health_score >= 80:
        status = "healthy"
    elif health_score >= 60:
        status = "warning"
    else:
        status = "critical"

    # Trend: compare with last score
    trend = await _calc_trend(session, device.id, health_score)

    score = DeviceHealthScore(
        device_id=device.id,
        timestamp=now,
        health_score=health_score,
        status=status,
        factors=factors,
        trend=trend,
    )
    session.add(score)
    return score


async def _calc_power_stability(session: AsyncSession, device_id: int, now: datetime) -> float:
    """Power stability score: low std_dev = high score."""
    cutoff = now - timedelta(hours=24)
    result = await session.execute(
        select(
            func.avg(EnergyData.value).label("avg_val"),
            func.stddev(EnergyData.value).label("std_val"),
        ).where(and_(
            EnergyData.device_id == device_id,
            EnergyData.data_type == "power",
            EnergyData.timestamp >= cutoff,
        ))
    )
    row = result.one_or_none()
    if not row or row.avg_val is None or row.avg_val == 0:
        return 85.0  # no data = assume OK
    avg_val = float(row.avg_val)
    std_val = float(row.std_val or 0)
    cv = std_val / avg_val if avg_val > 0 else 0  # coefficient of variation
    # cv < 0.1 = 100, cv > 0.5 = 40
    score = max(40, min(100, 100 - (cv - 0.1) * 150))
    return round(score, 1)


async def _calc_efficiency_score(session: AsyncSession, device: Device, now: datetime) -> float:
    """Efficiency score based on device type."""
    cutoff = now - timedelta(hours=24)
    if device.device_type == "heat_pump":
        result = await session.execute(
            select(func.avg(EnergyData.value)).where(and_(
                EnergyData.device_id == device.id,
                EnergyData.data_type == "cop",
                EnergyData.timestamp >= cutoff,
            ))
        )
        avg_cop = result.scalar()
        if avg_cop is None:
            return 85.0
        rated_cop = 3.5
        ratio = float(avg_cop) / rated_cop
        return round(max(30, min(100, ratio * 100)), 1)
    elif device.device_type == "pv_inverter":
        rated_power = device.rated_power or 110.0
        result = await session.execute(
            select(func.max(EnergyData.value)).where(and_(
                EnergyData.device_id == device.id,
                EnergyData.data_type == "power",
                EnergyData.timestamp >= cutoff,
            ))
        )
        max_power = result.scalar()
        if max_power is None:
            return 85.0
        ratio = float(max_power) / rated_power
        return round(max(30, min(100, ratio * 110)), 1)
    elif device.device_type == "meter":
        result = await session.execute(
            select(func.avg(EnergyData.value)).where(and_(
                EnergyData.device_id == device.id,
                EnergyData.data_type == "power_factor",
                EnergyData.timestamp >= cutoff,
            ))
        )
        avg_pf = result.scalar()
        if avg_pf is None:
            return 85.0
        return round(max(40, min(100, float(avg_pf) * 105)), 1)
    return 85.0


async def _calc_alarm_frequency_score(session: AsyncSession, device_id: int, now: datetime) -> float:
    """Fewer alarms = higher score. 0 alarms = 100, 10+ = 30."""
    cutoff = now - timedelta(days=7)
    result = await session.execute(
        select(func.count(AlarmEvent.id)).where(and_(
            AlarmEvent.device_id == device_id,
            AlarmEvent.triggered_at >= cutoff,
        ))
    )
    count = result.scalar() or 0
    score = max(30, 100 - count * 7)
    return float(score)


async def _calc_uptime_score(session: AsyncSession, device: Device, now: datetime) -> float:
    """Uptime based on device status and data availability."""
    cutoff = now - timedelta(days=7)
    result = await session.execute(
        select(func.count(EnergyData.id)).where(and_(
            EnergyData.device_id == device.id,
            EnergyData.data_type == "power",
            EnergyData.timestamp >= cutoff,
        ))
    )
    data_count = result.scalar() or 0
    # Expected ~4 readings/min * 60 * 24 * 7 = ~40320 (at 15s interval)
    expected = 7 * 24 * 60 * 4
    ratio = min(1.0, data_count / max(1, expected))
    # Also check current status
    status_penalty = 0
    if device.status == "offline":
        status_penalty = 15
    elif device.status == "alarm":
        status_penalty = 5
    return round(max(30, ratio * 100 - status_penalty), 1)


async def _calc_temperature_score(session: AsyncSession, device: Device, now: datetime) -> float:
    """Temperature within normal range = high score."""
    cfg = DEVICE_RATED_EFFICIENCY.get(device.device_type, {})
    temp_range = cfg.get("temp_range")
    if not temp_range:
        return 90.0  # N/A for this device type

    cutoff = now - timedelta(hours=6)
    result = await session.execute(
        select(func.avg(EnergyData.value)).where(and_(
            EnergyData.device_id == device.id,
            EnergyData.data_type == "temperature",
            EnergyData.timestamp >= cutoff,
        ))
    )
    avg_temp = result.scalar()
    if avg_temp is None:
        return 85.0
    avg_temp = float(avg_temp)
    low, high = temp_range
    if low <= avg_temp <= high:
        return 100.0
    deviation = max(0, avg_temp - high, low - avg_temp)
    return round(max(20, 100 - deviation * 5), 1)


async def _calc_trend(session: AsyncSession, device_id: int, current_score: float) -> str:
    """Compare with recent scores to determine trend."""
    result = await session.execute(
        select(DeviceHealthScore.health_score)
        .where(DeviceHealthScore.device_id == device_id)
        .order_by(DeviceHealthScore.timestamp.desc())
        .limit(5)
    )
    scores = [float(r) for r in result.scalars().all()]
    if len(scores) < 2:
        return "stable"
    avg_prev = sum(scores) / len(scores)
    diff = current_score - avg_prev
    if diff > 3:
        return "improving"
    elif diff < -3:
        return "degrading"
    return "stable"


# ── Anomaly Detection ───────────────────────────────────────────────

async def scan_anomalies(session: AsyncSession, device_id: int | None = None) -> list[AnomalyDetection]:
    """Scan for anomalies across devices using Z-score and pattern-based methods."""
    now = datetime.now(timezone.utc)
    anomalies = []

    query = select(Device).where(Device.is_active == True)
    if device_id:
        query = query.where(Device.id == device_id)
    result = await session.execute(query)
    devices = result.scalars().all()

    for device in devices:
        # Z-score based detection
        device_anomalies = await _zscore_detection(session, device, now)
        anomalies.extend(device_anomalies)

        # Pattern-based detection
        pattern_anomalies = await _pattern_detection(session, device, now)
        anomalies.extend(pattern_anomalies)

    for a in anomalies:
        session.add(a)

    return anomalies


async def _zscore_detection(
    session: AsyncSession, device: Device, now: datetime
) -> list[AnomalyDetection]:
    """Detect anomalies using Z-score method (> 3 sigma)."""
    anomalies = []
    metrics = ["power"]
    if device.device_type == "heat_pump":
        metrics.append("cop")
    if device.device_type == "sensor":
        metrics = ["temperature"]

    for metric in metrics:
        # Get rolling stats from last 24h
        cutoff_stats = now - timedelta(hours=24)
        stats_result = await session.execute(
            select(
                func.avg(EnergyData.value).label("avg_val"),
                func.stddev(EnergyData.value).label("std_val"),
            ).where(and_(
                EnergyData.device_id == device.id,
                EnergyData.data_type == metric,
                EnergyData.timestamp >= cutoff_stats,
            ))
        )
        stats = stats_result.one_or_none()
        if not stats or stats.avg_val is None or stats.std_val is None or float(stats.std_val) == 0:
            continue

        avg_val = float(stats.avg_val)
        std_val = float(stats.std_val)

        # Check latest value
        latest_result = await session.execute(
            select(EnergyData.value).where(and_(
                EnergyData.device_id == device.id,
                EnergyData.data_type == metric,
            )).order_by(EnergyData.timestamp.desc()).limit(1)
        )
        latest = latest_result.scalar()
        if latest is None:
            continue
        latest = float(latest)

        z_score = abs(latest - avg_val) / std_val
        if z_score > 3:
            deviation_pct = round(abs(latest - avg_val) / avg_val * 100, 1) if avg_val != 0 else 0
            severity = "critical" if z_score > 5 else "warning"
            anomaly_type = _classify_anomaly(device.device_type, metric, latest, avg_val)

            anomalies.append(AnomalyDetection(
                device_id=device.id,
                detected_at=now,
                anomaly_type=anomaly_type,
                severity=severity,
                description=f"{metric} 异常: 当前值 {latest:.2f}, 均值 {avg_val:.2f}, Z-score {z_score:.1f}",
                metric_name=metric,
                expected_value=round(avg_val, 2),
                actual_value=round(latest, 2),
                deviation_percent=deviation_pct,
                status="detected",
            ))

    return anomalies


def _classify_anomaly(device_type: str, metric: str, actual: float, expected: float) -> str:
    """Classify anomaly type based on device type and metric."""
    if metric == "power" and actual < expected:
        return "power_drop"
    if metric == "cop" and actual < expected:
        return "efficiency_loss"
    if metric == "temperature":
        return "abnormal_temperature"
    return "pattern_deviation"


async def _pattern_detection(
    session: AsyncSession, device: Device, now: datetime
) -> list[AnomalyDetection]:
    """Pattern-based anomaly detection for specific device types."""
    anomalies = []

    # Check for communication loss (no data in last 5 minutes)
    cutoff = now - timedelta(minutes=5)
    result = await session.execute(
        select(func.count(EnergyData.id)).where(and_(
            EnergyData.device_id == device.id,
            EnergyData.timestamp >= cutoff,
        ))
    )
    count = result.scalar() or 0
    if count == 0 and device.status == "online":
        anomalies.append(AnomalyDetection(
            device_id=device.id,
            detected_at=now,
            anomaly_type="communication_loss",
            severity="warning",
            description=f"设备 {device.name} 超过5分钟无数据上报",
            metric_name="data_availability",
            expected_value=1.0,
            actual_value=0.0,
            deviation_percent=100.0,
            status="detected",
        ))

    # PV specific: power drop during daytime (8:00-17:00 Beijing time)
    if device.device_type == "pv_inverter":
        beijing_hour = (now + timedelta(hours=8)).hour
        if 8 <= beijing_hour <= 17:
            latest_result = await session.execute(
                select(EnergyData.value).where(and_(
                    EnergyData.device_id == device.id,
                    EnergyData.data_type == "power",
                )).order_by(EnergyData.timestamp.desc()).limit(1)
            )
            power = latest_result.scalar()
            rated = device.rated_power or 110.0
            if power is not None and float(power) < rated * 0.05 and beijing_hour >= 9 and beijing_hour <= 16:
                anomalies.append(AnomalyDetection(
                    device_id=device.id,
                    detected_at=now,
                    anomaly_type="power_drop",
                    severity="warning",
                    description=f"光伏 {device.name} 在日照时段功率异常偏低: {power:.1f} kW",
                    metric_name="power",
                    expected_value=round(rated * 0.3, 2),
                    actual_value=round(float(power), 2),
                    deviation_percent=round((1 - float(power) / (rated * 0.3)) * 100, 1),
                    status="detected",
                ))

    # Heat pump: COP degradation
    if device.device_type == "heat_pump":
        latest_result = await session.execute(
            select(EnergyData.value).where(and_(
                EnergyData.device_id == device.id,
                EnergyData.data_type == "cop",
            )).order_by(EnergyData.timestamp.desc()).limit(1)
        )
        cop = latest_result.scalar()
        if cop is not None and float(cop) < 2.0:
            anomalies.append(AnomalyDetection(
                device_id=device.id,
                detected_at=now,
                anomaly_type="efficiency_loss",
                severity="warning" if float(cop) >= 1.5 else "critical",
                description=f"热泵 {device.name} COP降至 {cop:.2f}，低于正常水平",
                metric_name="cop",
                expected_value=3.5,
                actual_value=round(float(cop), 2),
                deviation_percent=round((1 - float(cop) / 3.5) * 100, 1),
                status="detected",
            ))

    return anomalies


# ── Diagnostic Intelligence ─────────────────────────────────────────

async def run_diagnostics(
    session: AsyncSession, device_id: int, report_type: str = "triggered"
) -> DiagnosticReport:
    """Run diagnostic analysis for a device and generate report."""
    now = datetime.now(timezone.utc)

    result = await session.execute(select(Device).where(Device.id == device_id))
    device = result.scalar_one_or_none()
    if not device:
        raise ValueError(f"Device {device_id} not found")

    findings = []
    recommendations = []
    energy_loss_kwh = 0.0
    cost_impact_yuan = 0.0

    # Check recent anomalies
    cutoff = now - timedelta(days=7)
    anomaly_result = await session.execute(
        select(AnomalyDetection).where(and_(
            AnomalyDetection.device_id == device_id,
            AnomalyDetection.detected_at >= cutoff,
        )).order_by(AnomalyDetection.detected_at.desc())
    )
    anomalies = anomaly_result.scalars().all()

    # Check alarm history
    alarm_result = await session.execute(
        select(AlarmEvent).where(and_(
            AlarmEvent.device_id == device_id,
            AlarmEvent.triggered_at >= cutoff,
        ))
    )
    alarms = alarm_result.scalars().all()

    # Get current health
    health_result = await session.execute(
        select(DeviceHealthScore).where(
            DeviceHealthScore.device_id == device_id
        ).order_by(DeviceHealthScore.timestamp.desc()).limit(1)
    )
    health = health_result.scalar_one_or_none()

    # Generate findings based on device type
    if device.device_type == "pv_inverter":
        findings, recommendations, energy_loss_kwh, cost_impact_yuan = \
            await _diagnose_pv(session, device, anomalies, alarms, health, now)
    elif device.device_type == "heat_pump":
        findings, recommendations, energy_loss_kwh, cost_impact_yuan = \
            await _diagnose_heat_pump(session, device, anomalies, alarms, health, now)
    else:
        findings, recommendations, energy_loss_kwh, cost_impact_yuan = \
            _diagnose_general(device, anomalies, alarms, health)

    report = DiagnosticReport(
        device_id=device_id,
        generated_at=now,
        report_type=report_type,
        findings=findings,
        recommendations=recommendations,
        estimated_impact={"energy_loss_kwh": round(energy_loss_kwh, 2), "cost_impact_yuan": round(cost_impact_yuan, 2)},
        status="generated",
    )
    session.add(report)
    return report


async def _diagnose_pv(session, device, anomalies, alarms, health, now):
    """Diagnostic logic for PV inverters."""
    findings = []
    recommendations = []
    energy_loss = 0.0
    cost_impact = 0.0

    # Check temperature
    cutoff = now - timedelta(hours=6)
    temp_result = await session.execute(
        select(func.avg(EnergyData.value)).where(and_(
            EnergyData.device_id == device.id,
            EnergyData.data_type == "temperature",
            EnergyData.timestamp >= cutoff,
        ))
    )
    avg_temp = temp_result.scalar()

    power_anomalies = [a for a in anomalies if a.anomaly_type == "power_drop"]
    temp_anomalies = [a for a in anomalies if a.anomaly_type == "abnormal_temperature"]

    if power_anomalies and avg_temp and float(avg_temp) > 55:
        findings.append({
            "finding": "光伏逆变器功率下降伴随高温",
            "severity": "warning",
            "detail": f"平均温度 {float(avg_temp):.1f}°C，高于正常范围。功率异常 {len(power_anomalies)} 次",
        })
        recommendations.append({
            "action": "清洁光伏面板并检查通风散热",
            "priority": "high",
            "detail": "高温导致逆变器降额运行，建议清洁面板、检查散热风扇",
        })
        energy_loss = len(power_anomalies) * 5.0
        cost_impact = energy_loss * 0.6

    elif power_anomalies:
        findings.append({
            "finding": "光伏功率间歇性下降",
            "severity": "info",
            "detail": f"近7天检测到 {len(power_anomalies)} 次功率下降",
        })
        recommendations.append({
            "action": "检查光伏面板遮挡和接线",
            "priority": "medium",
            "detail": "排除树木遮挡、面板污染或接线松动",
        })

    if alarms:
        findings.append({
            "finding": f"近7天触发 {len(alarms)} 条告警",
            "severity": "warning" if len(alarms) > 3 else "info",
            "detail": "频繁告警可能指示设备劣化",
        })

    if health and health.health_score < 70:
        findings.append({
            "finding": f"设备健康评分偏低: {health.health_score}",
            "severity": "warning",
            "detail": f"当前趋势: {health.trend}",
        })
        recommendations.append({
            "action": "安排专项巡检",
            "priority": "high",
            "detail": "建议安排技术人员进行全面检查",
        })

    if not findings:
        findings.append({
            "finding": "设备运行状态良好",
            "severity": "info",
            "detail": "未发现明显异常",
        })

    return findings, recommendations, energy_loss, cost_impact


async def _diagnose_heat_pump(session, device, anomalies, alarms, health, now):
    """Diagnostic logic for heat pumps."""
    findings = []
    recommendations = []
    energy_loss = 0.0
    cost_impact = 0.0

    # Check outdoor temperature for context
    cutoff = now - timedelta(hours=6)
    cop_anomalies = [a for a in anomalies if a.anomaly_type == "efficiency_loss"]

    # Check outdoor temp via sensors
    outdoor_result = await session.execute(
        select(func.avg(EnergyData.value)).where(and_(
            EnergyData.data_type == "outdoor_temp",
            EnergyData.timestamp >= cutoff,
        ))
    )
    outdoor_temp = outdoor_result.scalar()

    if cop_anomalies and outdoor_temp and float(outdoor_temp) < -5:
        findings.append({
            "finding": "热泵COP下降，与低温天气相关",
            "severity": "info",
            "detail": f"室外温度 {float(outdoor_temp):.1f}°C，COP降额属正常现象",
        })
        recommendations.append({
            "action": "调整运行策略以适应低温",
            "priority": "low",
            "detail": "低温环境下COP降低属正常特性，可适当调整运行时段",
        })
    elif cop_anomalies:
        findings.append({
            "finding": "热泵能效异常下降",
            "severity": "warning",
            "detail": f"检测到 {len(cop_anomalies)} 次COP异常",
        })
        recommendations.append({
            "action": "检查冷媒充注量和换热器",
            "priority": "high",
            "detail": "COP异常降低可能由冷媒泄漏或换热器结垢导致",
        })
        energy_loss = len(cop_anomalies) * 8.0
        cost_impact = energy_loss * 0.6

    comm_anomalies = [a for a in anomalies if a.anomaly_type == "communication_loss"]
    if comm_anomalies:
        findings.append({
            "finding": f"通讯中断 {len(comm_anomalies)} 次",
            "severity": "warning",
            "detail": "频繁通讯丢失需检查网络和DTU设备",
        })
        recommendations.append({
            "action": "检查通讯网络和DTU",
            "priority": "medium",
            "detail": "检查DTU设备状态、网线连接和信号强度",
        })

    if health and health.health_score < 70:
        findings.append({
            "finding": f"设备健康评分偏低: {health.health_score}",
            "severity": "warning",
            "detail": f"当前趋势: {health.trend}",
        })

    if not findings:
        findings.append({
            "finding": "设备运行状态良好",
            "severity": "info",
            "detail": "未发现明显异常",
        })

    return findings, recommendations, energy_loss, cost_impact


def _diagnose_general(device, anomalies, alarms, health):
    """Generic diagnostic for other device types."""
    findings = []
    recommendations = []
    energy_loss = 0.0
    cost_impact = 0.0

    if anomalies:
        by_type = {}
        for a in anomalies:
            by_type.setdefault(a.anomaly_type, []).append(a)
        for atype, items in by_type.items():
            findings.append({
                "finding": f"检测到 {len(items)} 次 {atype} 异常",
                "severity": "warning" if len(items) > 2 else "info",
                "detail": items[0].description if items else "",
            })

    if alarms:
        findings.append({
            "finding": f"近7天触发 {len(alarms)} 条告警",
            "severity": "warning" if len(alarms) > 3 else "info",
            "detail": "频繁告警需要关注",
        })

    if health and health.health_score < 70:
        recommendations.append({
            "action": "安排设备巡检",
            "priority": "high",
            "detail": f"健康评分 {health.health_score}，趋势 {health.trend}",
        })

    if not findings:
        findings.append({
            "finding": "设备运行正常",
            "severity": "info",
            "detail": "未发现异常",
        })

    return findings, recommendations, energy_loss, cost_impact


# ── Predictive Maintenance ──────────────────────────────────────────

async def generate_maintenance_predictions(session: AsyncSession) -> list[MaintenancePrediction]:
    """Generate maintenance predictions based on health trends and patterns."""
    now = datetime.now(timezone.utc)
    predictions = []

    result = await session.execute(select(Device).where(Device.is_active == True))
    devices = result.scalars().all()

    for device in devices:
        # Get recent health scores
        health_result = await session.execute(
            select(DeviceHealthScore).where(
                DeviceHealthScore.device_id == device.id
            ).order_by(DeviceHealthScore.timestamp.desc()).limit(10)
        )
        scores = health_result.scalars().all()

        if not scores:
            continue

        latest = scores[0]

        # Rule: health score < 60 and degrading
        if latest.health_score < 60 and latest.trend == "degrading":
            days_to_failure = max(3, int((latest.health_score - 20) / 5))
            predictions.append(MaintenancePrediction(
                device_id=device.id,
                predicted_at=now,
                component=_get_weak_component(latest.factors),
                failure_mode="设备性能持续下降，可能导致故障停机",
                probability=round(min(0.9, (100 - latest.health_score) / 100 + 0.2), 2),
                predicted_failure_date=now + timedelta(days=days_to_failure),
                recommended_action="安排全面检修，重点检查薄弱环节",
                urgency="critical" if latest.health_score < 40 else "high",
                estimated_downtime_hours=4.0 if device.device_type in ("heat_pump", "pv_inverter") else 2.0,
                estimated_repair_cost=_estimate_repair_cost(device.device_type),
                status="predicted",
            ))

        # Rule: health between 60-75 and degrading
        elif 60 <= latest.health_score < 75 and latest.trend == "degrading":
            predictions.append(MaintenancePrediction(
                device_id=device.id,
                predicted_at=now,
                component=_get_weak_component(latest.factors),
                failure_mode="性能下降趋势，需预防性维护",
                probability=round(min(0.6, (80 - latest.health_score) / 100 + 0.1), 2),
                predicted_failure_date=now + timedelta(days=14),
                recommended_action="安排预防性巡检和维护",
                urgency="medium",
                estimated_downtime_hours=2.0,
                estimated_repair_cost=_estimate_repair_cost(device.device_type) * 0.5,
                status="predicted",
            ))

        # Rule: check alarm frequency
        alarm_score = latest.factors.get("alarm_frequency", 100) if latest.factors else 100
        if alarm_score < 50:
            predictions.append(MaintenancePrediction(
                device_id=device.id,
                predicted_at=now,
                component="告警系统/传感器",
                failure_mode="频繁告警可能预示设备故障",
                probability=0.4,
                predicted_failure_date=now + timedelta(days=7),
                recommended_action="排查告警根因，检查传感器和控制系统",
                urgency="medium",
                estimated_downtime_hours=1.0,
                estimated_repair_cost=500.0,
                status="predicted",
            ))

    for p in predictions:
        session.add(p)
    return predictions


def _get_weak_component(factors: dict | None) -> str:
    """Identify the weakest factor as the component needing attention."""
    if not factors:
        return "综合"
    component_map = {
        "power_stability": "电力输出系统",
        "efficiency": "能效/换热系统",
        "alarm_frequency": "监控传感器",
        "uptime": "通讯/控制系统",
        "temperature": "散热/温控系统",
    }
    weakest = min(factors, key=lambda k: factors.get(k, 100))
    return component_map.get(weakest, "综合")


def _estimate_repair_cost(device_type: str) -> float:
    """Estimate repair cost by device type."""
    costs = {
        "pv_inverter": 3000.0,
        "heat_pump": 5000.0,
        "meter": 800.0,
        "sensor": 300.0,
        "heat_meter": 1500.0,
    }
    return costs.get(device_type, 1000.0)


# ── Operational Insights ────────────────────────────────────────────

async def generate_insights(session: AsyncSession) -> list[OpsInsight]:
    """Generate operational insights from data analysis."""
    now = datetime.now(timezone.utc)
    insights = []

    # Insight 1: Device efficiency comparison
    result = await session.execute(
        select(
            DeviceHealthScore.device_id,
            func.avg(DeviceHealthScore.health_score).label("avg_score"),
        ).where(
            DeviceHealthScore.timestamp >= now - timedelta(days=7)
        ).group_by(DeviceHealthScore.device_id)
    )
    scores_by_device = result.all()

    if scores_by_device:
        scores = [float(s.avg_score) for s in scores_by_device]
        avg_all = sum(scores) / len(scores) if scores else 0
        low_performers = [s for s in scores_by_device if float(s.avg_score) < avg_all - 10]

        if low_performers:
            device_ids = [s.device_id for s in low_performers]
            dev_result = await session.execute(
                select(Device.id, Device.name).where(Device.id.in_(device_ids))
            )
            device_names = {r.id: r.name for r in dev_result.all()}
            insights.append(OpsInsight(
                insight_type="performance_comparison",
                title="部分设备健康评分低于平均水平",
                description=f"以下设备健康评分低于园区平均值({avg_all:.0f})超过10分: "
                    + ", ".join(device_names.get(d, f"#{d}") for d in device_ids),
                data={"avg_score": round(avg_all, 1), "low_performers": [
                    {"device_id": s.device_id, "score": round(float(s.avg_score), 1),
                     "name": device_names.get(s.device_id, f"#{s.device_id}")}
                    for s in low_performers
                ]},
                impact_level="medium" if len(low_performers) <= 2 else "high",
                actionable=True,
                recommended_action="重点关注低评分设备，安排巡检和维护",
                generated_at=now,
                valid_until=now + timedelta(days=7),
            ))

    # Insight 2: Anomaly trend
    week_ago = now - timedelta(days=7)
    two_weeks_ago = now - timedelta(days=14)

    this_week = await session.execute(
        select(func.count(AnomalyDetection.id)).where(
            AnomalyDetection.detected_at >= week_ago
        )
    )
    last_week = await session.execute(
        select(func.count(AnomalyDetection.id)).where(and_(
            AnomalyDetection.detected_at >= two_weeks_ago,
            AnomalyDetection.detected_at < week_ago,
        ))
    )
    this_count = this_week.scalar() or 0
    last_count = last_week.scalar() or 0

    if this_count > last_count * 1.5 and this_count > 3:
        insights.append(OpsInsight(
            insight_type="efficiency_trend",
            title="异常检测数量环比上升",
            description=f"本周检测到 {this_count} 次异常，上周 {last_count} 次，增长 {((this_count/max(1,last_count))-1)*100:.0f}%",
            data={"this_week": this_count, "last_week": last_count},
            impact_level="high",
            actionable=True,
            recommended_action="建议全面排查设备状态，加强巡检频次",
            generated_at=now,
            valid_until=now + timedelta(days=3),
        ))

    # Insight 3: Energy cost optimization
    insights.append(OpsInsight(
        insight_type="cost_anomaly",
        title="能源使用效率周报",
        description="园区整体运行状况评估",
        data={
            "total_devices": len(scores_by_device) if scores_by_device else 0,
            "avg_health": round(avg_all, 1) if scores_by_device else 0,
            "anomaly_count": this_count,
        },
        impact_level="low",
        actionable=False,
        generated_at=now,
        valid_until=now + timedelta(days=7),
    ))

    for i in insights:
        session.add(i)
    return insights


# ── Dashboard Aggregation ───────────────────────────────────────────

async def get_dashboard_data(session: AsyncSession) -> dict:
    """Get AI Ops dashboard overview data."""
    now = datetime.now(timezone.utc)

    # Latest health scores per device
    subq = (
        select(
            DeviceHealthScore.device_id,
            func.max(DeviceHealthScore.timestamp).label("max_ts"),
        ).group_by(DeviceHealthScore.device_id).subquery()
    )
    health_result = await session.execute(
        select(DeviceHealthScore).join(
            subq, and_(
                DeviceHealthScore.device_id == subq.c.device_id,
                DeviceHealthScore.timestamp == subq.c.max_ts,
            )
        )
    )
    health_scores = health_result.scalars().all()

    # Get device names
    device_ids = [h.device_id for h in health_scores]
    if device_ids:
        dev_result = await session.execute(
            select(Device.id, Device.name, Device.device_type).where(Device.id.in_(device_ids))
        )
        device_map = {r.id: {"name": r.name, "type": r.device_type} for r in dev_result.all()}
    else:
        device_map = {}

    health_summary = {
        "healthy": sum(1 for h in health_scores if h.status == "healthy"),
        "warning": sum(1 for h in health_scores if h.status == "warning"),
        "critical": sum(1 for h in health_scores if h.status == "critical"),
        "avg_score": round(sum(h.health_score for h in health_scores) / max(1, len(health_scores)), 1),
        "devices": [{
            "device_id": h.device_id,
            "device_name": device_map.get(h.device_id, {}).get("name", f"#{h.device_id}"),
            "device_type": device_map.get(h.device_id, {}).get("type", "unknown"),
            "health_score": h.health_score,
            "status": h.status,
            "trend": h.trend,
            "factors": h.factors,
        } for h in health_scores],
    }

    # Recent anomalies
    anomaly_result = await session.execute(
        select(AnomalyDetection)
        .where(AnomalyDetection.detected_at >= now - timedelta(days=7))
        .order_by(AnomalyDetection.detected_at.desc())
        .limit(20)
    )
    recent_anomalies = anomaly_result.scalars().all()
    anomaly_stats = {
        "total": len(recent_anomalies),
        "by_severity": {},
        "by_type": {},
    }
    for a in recent_anomalies:
        anomaly_stats["by_severity"][a.severity] = anomaly_stats["by_severity"].get(a.severity, 0) + 1
        anomaly_stats["by_type"][a.anomaly_type] = anomaly_stats["by_type"].get(a.anomaly_type, 0) + 1

    # Maintenance predictions
    pred_result = await session.execute(
        select(MaintenancePrediction).where(
            MaintenancePrediction.status == "predicted"
        ).order_by(MaintenancePrediction.urgency.desc()).limit(10)
    )
    predictions = pred_result.scalars().all()

    # Latest insights
    insight_result = await session.execute(
        select(OpsInsight).where(
            OpsInsight.valid_until >= now
        ).order_by(OpsInsight.generated_at.desc()).limit(5)
    )
    latest_insights = insight_result.scalars().all()

    return {
        "health": health_summary,
        "anomalies": {
            "stats": anomaly_stats,
            "recent": [{
                "id": a.id,
                "device_id": a.device_id,
                "device_name": device_map.get(a.device_id, {}).get("name", f"#{a.device_id}"),
                "anomaly_type": a.anomaly_type,
                "severity": a.severity,
                "description": a.description,
                "detected_at": str(a.detected_at),
                "status": a.status,
            } for a in recent_anomalies[:10]],
        },
        "predictions": [{
            "id": p.id,
            "device_id": p.device_id,
            "device_name": device_map.get(p.device_id, {}).get("name", f"#{p.device_id}"),
            "component": p.component,
            "failure_mode": p.failure_mode,
            "probability": p.probability,
            "predicted_failure_date": str(p.predicted_failure_date) if p.predicted_failure_date else None,
            "urgency": p.urgency,
            "recommended_action": p.recommended_action,
        } for p in predictions],
        "insights": [{
            "id": i.id,
            "insight_type": i.insight_type,
            "title": i.title,
            "description": i.description,
            "impact_level": i.impact_level,
            "actionable": i.actionable,
            "recommended_action": i.recommended_action,
            "generated_at": str(i.generated_at),
        } for i in latest_insights],
    }